AI product designer & developer · Texas, USA
Prasiddha K.AI systems builder.
I build the evidence, tools, interfaces, and safeguards that trustworthy AI products need.
The ten working products below demonstrate public-data retrieval, revenue automation, transparent logic, multimodal interaction, business-case modeling, evaluation, and human control. Every live system states its AI boundary.
- 10
- Working products
- 07
- Public-data systems
- 03
- Browser-local tools
Portfolio in 93 seconds
Ten products.One responsible AI system philosophy.
A concise guided tour of how evidence becomes a useful result—through bounded logic, visible uncertainty, human review, and verification.
Natural neural narrationOpen captionsOriginal score + sound designUser-controlled playback
About the builder
Systems thinking,
made tangible.
I’m Prasiddha Karki, a Texas-based AI product designer and developer building end-to-end workflows around public data, retrieval, transparent rules, evaluation, and human control.
The work here is intentionally inspectable: each product shows its sources, deterministic boundaries, known limitations, and the decision that still belongs to a person. I’m also an MBA and MS Data Science candidate at Eastern University, bringing experience in insurance and healthcare administration. I’m available for focused product design, prototypes, and AI system collaboration with teams working globally.
- 10
- Working products
- 169
- Automated release checks
- 100%
- AI boundary disclosed
Start with your question
What do you need to investigate?
Choose a goal and jump directly into the matching live product or browser-local tool.
/01Weather & air quality
Plan around environmental conditions
Search any city or postal code for current weather, a simple good-or-caution verdict, and the nearest live city-level IQAir AQI observation.
Open HeatSignal ↓/02Electricity by Texas, state, or two ZIPsCompare source-backed electricity context
Inspect Texas hourly signals, explore monthly state benchmarks, or compare two ZIP-linked usage estimates and potential utility contexts with every proxy labeled.
Open Voltline ↓/03Vehicle safety & historyResearch recalls, title and history
Decode an authorized VIN, review matching campaigns, and open trustworthy title, salvage, owner, and accident-history checks.
Open FieldLens ↓/04Website due diligenceInspect a public website before trusting it
Enter a website name, domain, or URL to review available page evidence, registration context, and explainable risk signals.
Open SceneCraft ↓/05Food nearbyFind a restaurant for your craving
Search a city or neighborhood by flavor, cuisine, or dietary preference, then compare distance, menus, price signals, and map links.
Open PlateScout ↓/06Edit or sign a PDFAdd text, signatures and images privately
Open a PDF locally, place visible edits, review each page, and download the result without sending the document to a server.
Open PaperPatch ↓/07Automation business caseEstimate the value of repetitive work
Model saved hours, net monthly and annual value, subscription payback, and estimated ROI from your own assumptions.
Open ROI Calculator ↓/08Healthcare account qualificationScore and route healthcare organizations
Inspect an organization-level ICP score, buying-committee roles, qualified-account dashboard, and a reviewable HubSpot handoff.
Open CareSignal ↓/09Inbound sales orchestrationTurn an inquiry into an owned next step
Model an inbound inquiry with transparent scope, workload, integration, risk, region, and budget assumptions—without submitting the lead.
Open RoutePilot ↓/10Account researchBuild a source-aware sales brief
Generate a structured account brief with pain-point hypotheses, role targets, buying signals, questions, outreach, and confidence warnings.
Open AccountBrief ↓Live AI system / HeatSignal
Search anywhere.
Read the risk now.
A working environmental-intelligence demo: resolve any searchable location, pull current weather and the nearest live city-level IQAir US AQI observation, calculate heat risk, and inspect the source location on an interactive map.
Elapsed time from the first relevant condition check to a documented schedule, control, or escalation decision.
People responsible for field work can spot environmental concerns earlier and reach official guidance with less search friction.
Live AQI + current weather workflowIQAir observation · Open-Meteo weather
01Locate
02Weather
03Air quality
04Classify
05Cite
HEATSIGNAL / READY
Search cities and postal codes worldwide to combine current weather with the nearest live, city-level IQAir US AQI observation.
What I build
AI systems,
end to end.
A useful AI product is more than a model call. It needs reliable context, appropriate tools, clear permissions, measurable quality, and a thoughtful role for human judgment.
Retrieval & knowledge
Grounded answers over documents and structured data—with ingestion, hybrid search, citations, permissions, and freshness designed into the experience.
Agentic workflow design
A practice direction: planning tool use, visible state, scoped permissions, approval gates, and fallbacks. The public work below demonstrates the non-generative foundations—not a deployed autonomous agent.
Multimodal AI
Interfaces that combine text, images, audio, and structured inputs so people can work naturally across messy real-world information.
Evals & safety
Task-based eval sets, tracing, failure analysis, guardrails, and review queues that reveal when a system is uncertain or needs a person.
Ways I can help
From uncertain idea
to working evidence.
Choose the smallest engagement that answers the next important product question. Every path leaves you with visible decisions, boundaries, and a practical next step.
AI product diagnostic
For teams deciding what should be automated, what must stay human, and how quality will be measured.
DeliverableWorkflow map · source and risk review · human gates · practical evaluation plan
Discuss this path ↘Prototype sprint
For a real product question that needs a designed, working end-to-end vertical slice.
DeliverableExperience design · system flow · working prototype · evidence and failure-state review
Discuss this path ↘System + evaluation partnership
For teams improving an existing AI workflow, its interface, reliability, safeguards, or evaluation practice.
DeliverableProduct iteration · eval design · trace review · monitoring and human escalation
Discuss this path ↘Independent product lab
Selected AI work
Ten working products that turn live evidence, local files, business assumptions, or synthetic workflow fixtures into useful outcomes—without hiding the source, formula, privacy boundary, limitations, or role of human judgment.
Flagship case study / 01
Voltline
See the Texas hourly architecture, nationwide state layer, Census and ACS ZIP model, EIA-861 utility joins, comparison rules, and explicit evidence boundaries.
Read the build story ↗Flagship case study / 02SceneCraft
Explore the protected retrieval flow, partial-audit design, trust-signal ruleset, and tested failure handling behind the website preflight.
Read the build story ↗New portfolio collection
Revenue AI systems
Three end-to-end revenue workflow products with inspectable logic, official public research where coverage exists, synthetic evaluation fixtures, diagrams, walkthrough videos, technology notes, and production integration boundaries.
/08 · Healthcare GTM
CareSignal
Explainable 0–100 account scoring, role targets, qualification dashboard, and HubSpot-ready preview.
Open live project ↗/09 · Revenue operationsRoutePilot
Inquiry classification, implementation range, representative assignment, response draft, alert, and SLA trace.
Open live project ↗/10 · Account intelligenceAccountBrief
Sourced public facts, freshness warnings, operational hypotheses, discovery questions, outreach, and safe abstention.
Open live project ↗Interactive intermission / Browser game
Take the controls.
Defend the orbit.
Blast asteroid fields, survive surprise alien formations, clear escalating sectors, and face a different animal commander as the boss. It runs entirely in your browser and stays separate from every portfolio product.
- 01
- Instant play
- 02
- Optional sound FX
- 03
- Keyboard + touch
Live public dataIndependent exploration
/01Voltline
Texas grid signals plus nationwide ZIP-linked electricity context
Voltline combines Texas EIA-930 signals and monthly state benchmarks with 2020 Census ZCTA mapping, 2024 ACS occupied-housing estimates, and final 2024 EIA-861 utility context for one or two ZIP Codes.
Which locations need deeper electricity-cost or service-territory diligence before a facilities decision?
Elapsed time from an approved location-comparison request to a human-reviewed shortlist ready for utility verification.
Finance and facilities teams get a comparable starting point without manually joining incompatible geography and energy datasets.
01Enter one or two ZIPs
02Map Census geography
03Join EIA evidence
04Compare with boundaries
System disclosure · AI boundary · evidence
System typeDeterministic public-data and geographic-context system
AI boundaryGenerative model: none. Fixed rules align public signals and calculate an explicitly modeled ZIP-area range from public aggregates.
Human gateThe visitor verifies service, rates, offers, bills, and operational meaning with the utility, regulator, EIA, or ERCOT; Voltline takes no action.
Current evidenceLive-source wiring, versioned Census/ACS/EIA-861 references, timestamps, missing-value handling, component labels, cache behavior, and fallback UI are implemented.
Latest public data · U.S. Energy Information Administration
Texas hourly + state + ZIP-linked contextU.S. electricity context
01Query EIA-930
02Normalize signals
03Align forecast
04Explain variance
Load the latest available EIA-930 data for Texas, or scroll this monitor into view to connect automatically.
Method, safeguards & evaluation
How the system works
A cached read-only service aligns hourly Texas evidence, resolves a ZIP-shaped Census geography to county and state, models area usage from occupied housing and a state residential benchmark, and joins county-level utility candidates without calling any result measured ZIP usage or a current rate.
Human control & safety
The interface never calls a public-data pattern a reliability alert, measured ZIP consumption, a confirmed utility, a tariff, or a supplier offer. Every reporting period, proxy, and estimate remains visible.
Evaluation plan
Source freshness · ZCTA integrity · estimate math · utility-component separation
Live public dataIndependent exploration
/02HeatSignal
Live environmental intelligence for heat-aware decisions anywhere
HeatSignal turns a city, region, or postal code into a current, source-backed snapshot of temperature, heat index, the nearest live city-level IQAir US AQI observation, and a plain-language outdoor-conditions verdict.
01Resolve the place
02Fetch weather + live AQI
03Classify conditions
04Map + guide
System disclosure · AI boundary · evidence
System typeRules-based environmental decision support
AI boundaryGenerative model: none. Current modeled weather, live city-level IQAir AQI, the NWS heat-index formula, and fixed risk bands produce the result.
Human gateVisitors can correct an ambiguous place match and are directed to official weather, air-quality, and local guidance before acting.
Current evidenceIQAir observation timestamps, protected-key behavior, partial air-quality handling, AirNow color bands, and the plain-language condition verdict are implemented and covered by release checks.
Method, safeguards & evaluation
How the system works
A bounded server-side integration protects the IQAir API key while the client geocodes the search, retrieves current weather and the nearest city-level live AQI in parallel, calculates the NWS heat index when conditions meet its threshold, applies AirNow AQI color bands, and keeps observation timestamps and source boundaries visible.
Human control & safety
City-level IQAir observations are not presented as a specific neighborhood station. Missing values remain visible without a modeled substitute, no personal health outcome is predicted, and official NWS, IQAir, and local guidance stays one click away.
Evaluation plan
API integrity · location accuracy · calculation tests · failure recovery
Live public dataIndependent exploration
/03FieldLens
A vehicle safety, recall, title and history research assistant
FieldLens turns a 17-character VIN into a readable vehicle identity, matching federal recall campaigns, urgent Park It or Park Outside flags, and trustworthy pathways for title, salvage, ownership, and accident-history checks.
Should an authorized vehicle move forward, be escalated, or be held for VIN-specific safety and title verification?
Elapsed time from authorized VIN submission to a documented human disposition with required follow-up identified.
Drivers, buyers, and fleet reviewers can find urgent campaign evidence and the correct history-check pathway without confusing a decode with a clean vehicle.
01Validate the VIN
02Decode vehicle
03Retrieve campaigns
04Route history checks
System disclosure · AI boundary · evidence
System typeFederal-record retrieval workflow
AI boundaryGenerative model: none. VIN normalization, source retrieval, and urgent-campaign ordering are deterministic.
Human gateThe visitor confirms authorization, VIN-specific recall status, and any title or history report with the responsible official provider.
Current evidenceAutomated checks cover VIN normalization and check digits, NHTSA decoding, recall-date parsing, partial-provider recovery, session-only notes, and prohibited safety claims.
Live public data · National Highway Traffic Safety Administration
No VIN is stored hereVIN safety, recall and history research assistant
01Validate VIN
02Decode vehicle
03Retrieve campaigns
04Route history checks
Enter a VIN to turn NHTSA identity data and matching recall campaigns into a readable safety brief.
Method, safeguards & evaluation
How the system works
A staged workflow validates and decodes the VIN through NHTSA vPIC, retrieves federal recall campaigns by year, make, and model, and separates those public safety records from title and history checks that require DOJ-approved NMVTIS or licensed providers.
Human control & safety
A successful VIN decode is never called a clean title. Unchecked title, owner, and accident fields stay visibly unknown; report-derived notes are labeled as user-entered, kept only for the page session, and never upgraded into a safety guarantee.
Evaluation plan
VIN validation · campaign recall · warning fidelity · history-source provenance
Live public dataIndependent exploration
/04SceneCraft
A live technical preflight and website risk-signal auditor
SceneCraft accepts a website name, domain, or URL and separates available technical readiness, accountability hints, external references, RDAP context, header posture, and common risk signals into explainable results—even when page HTML is limited.
Should a public website proceed to deeper vendor, customer, or partner due diligence—and which checks should happen next?
Elapsed time from approved public-website submission to a documented next-check or escalation decision.
Reviewers get a structured preflight that separates available signals from missing evidence instead of starting with an unbounded web search.
01Resolve + fetch page
02Check domain + risk
03Score available evidence
04Prioritize next checks
System disclosure · AI boundary · evidence
System typeDeterministic website preflight and risk-signal ruleset
AI boundaryGenerative model: none. Bounded retrieval, explicit scoring weights, regexes, and availability rules produce the audit.
Human gateVisitors confirm interpreted addresses, reveal high-attention results deliberately, and complete the linked manual checks before making a trust decision.
Current evidenceMocked route tests cover full and partial inspections, www fallback, private redirects, bot challenges, timeouts, and XHTML. Trust-score accuracy is not claimed.
Live public website inspection · no login required
HTML when available + domain signalsWebsite readiness, trust and risk audit
01Resolve + fetch page
02Check domain + risk
03Score available evidence
04Prioritize next checks
Enter a website name, domain, or public URL. SceneCraft inspects page content when available and still returns clearly labeled domain-level evidence when automated HTML access is limited.
Method, safeguards & evaluation
How the system works
A protected server workflow resolves simple .com names, tries the same registered domain’s www host when appropriate, verifies public DNS on every hop, streams bounded public HTML when available, and still returns clearly labeled RDAP, malware-filter DNS, and MDN header evidence when bot protection blocks page inspection.
Human control & safety
Missing HTML is never scored as a failure or a pass. Partial audits stay visibly partial, inferred .com addresses are shown for confirmation, and positive transparency never cancels a high-impact warning.
Evaluation plan
False-positive rate · coverage honesty · signal precision · explanation usefulness
Live public dataIndependent exploration
/05PlateScout
A craving-first nearby restaurant discovery assistant
PlateScout searches live public place data around any city, neighborhood, or postal code, then ranks restaurants by declared cuisine or dietary fit, distance, and the completeness of their menu, hours, and service information.
Which nearby meal options deserve confirmation for a team, traveler, or field crew’s stated location and constraints?
Elapsed time from a confirmed area and stated preference to a human-confirmed option.
People spend less time repeating nearby searches and get direct paths to verify menus, hours, directions, and dietary fit.
01Resolve the area
02Scan nearby food
03Rank transparent fit
04Map the shortlist
System disclosure · AI boundary · evidence
System typeGeospatial retrieval and heuristic ranking
AI boundaryGenerative model: none. Declared cuisine tags, distance, and source-data completeness drive a fixed, explainable rank.
Human gateVisitors verify menus, reviews, hours, preparation, and directions before traveling or spending money.
Current evidenceAutomated route tests cover non-Latin location isolation, explicit dietary non-matches, and street-address rejection before provider access. Relevance benchmarking remains planned.
Live nearby discovery · source-backed public place data
Craving + distance + menu links + mapFind food that fits your craving
01Locate
02Scan nearby places
03Match the craving
04Map the shortlist
PLATESCOUT / READY
Try “spicy” in Austin, a cuisine near your hotel, or a dietary preference in your neighborhood.
Method, safeguards & evaluation
How the system works
A server workflow geocodes the chosen area with Nominatim, queries nearby food places through Overpass, normalizes public OpenStreetMap tags, and applies transparent craving, distance, and data-completeness scoring without inventing ratings, popularity, or prices.
Human control & safety
Missing menu, price, popularity, and opening-hour data stays visibly missing. Visitors get direct menu, review, map, and direction links so they can confirm current information before traveling or spending money.
Evaluation plan
Location accuracy · cuisine precision · source coverage · graceful fallback
Live browser toolIndependent exploration
/06PaperPatch
A private, browser-local PDF editing and signing workspace
PaperPatch lets anyone open and quickly scan a PDF in their browser, review extractable-text and reading metrics, inspect a clearly labeled AI-like versus human-like writing-style estimate, add visible edits, and download the result without uploading the document to a server.
Can an eligible PDF be completed with visible local edits, or does it require a secure redaction, identity, signature, or document-system workflow?
Elapsed time from opening an approved, eligible PDF to a reviewed exported document.
People can complete ordinary visible PDF edits without uploading the file while being routed away from unsupported high-assurance tasks.
01Open locally
02Scan text layer
03Review signals and pages
04Edit and download
System disclosure · AI boundary · evidence
System typeBrowser-local document editor and exporter
AI boundaryGenerative model: none. The local scan uses bounded surface-pattern heuristics and intentionally limits the displayed signal; it cannot determine who or what authored a document.
Human gateThe visitor decides whether the extracted-text metrics are usable, reviews the original pages, and treats the style estimate only as a prompt for further review before editing or exporting.
Current evidenceAutomated checks cover PDF save/reopen behavior, scan sampling, insufficient-text handling, bounded scores, keyboard recovery, and modal focus. Cross-browser export-fidelity testing remains planned.
Private browser workspace · no account
Nothing is uploadedPrivate PDF editor
Drop a PDF here or choose a file
Your document, quick-scan text, images, signatures, and edits stay in this browser tab. Maximum 25 MB and 150 pages.
Choose a PDF to begin. Nothing is uploaded.
Know what this editor does
PaperPatch adds new text, drawings, images, highlights, and visual signatures. Its quick scan reads embedded text without OCR; the AI-like and human-like percentages describe writing patterns and cannot prove authorship. It does not rewrite arbitrary existing PDF text. A visual signature is not identity verification or a certificate-backed digital signature. A white cover-up is not secure redaction, and editing a previously signed PDF can invalidate its existing cryptographic signature. Password-protected PDFs may not open.
Measure a faster document workflow without giving up privacy.
Use PaperPatch first on authorized, low-risk documents. Compare the same document type with today's process before deciding whether it belongs in a larger operational workflow.
Proceed only when visible annotations and a downloaded copy meet the real review requirement.
Measure from the same start and finish points in the current workflow and the pilot.
Faster exports do not count as improvement if reviewers find more placement errors, missing pages, or unauthorized edits.
Baseline the current process, pilot only authorized files, record rework, and set a target after the baseline is known.
Stop condition: do not expand the pilot when the workflow requires secure redaction, identity verification, certificate-backed signatures, retained audit logs, or centralized records. PaperPatch does not provide those controls.
Method, safeguards & evaluation
How the system works
PDF.js renders the selected document and extracts its embedded text layer inside the visitor’s browser. A bounded, deterministic scan reports page coverage, word count, reading time, and surface writing-pattern signals; PaperPatch then keeps edits locally and uses pdf-lib to export a downloadable PDF.
Human control & safety
Files never leave the browser tab. The writing-style percentages are non-forensic pattern estimates—not an AI detector, plagiarism check, or proof of authorship—and image-only documents need separate OCR. PaperPatch does not verify identity, create certificate-backed signatures, or turn a white cover-up into secure redaction.
Evaluation plan
Text extraction · scan latency · placement accuracy · export fidelity · browser memory limits
Live browser toolIndependent exploration
/07Automation ROI
A transparent AI automation business-case calculator
The AI Automation ROI Calculator turns staffing, labor-cost, repetitive-work, automation, and software-cost assumptions into an inspectable estimate of saved capacity, net savings, subscription payback, and annual ROI.
Which repetitive workflow has enough measured capacity value to justify a controlled automation pilot?
Value of verified capacity the organization can actually reuse, less recurring automation cost.
Operators can compare automation ideas with visible economics instead of relying on promised savings or headline productivity claims.
01Define baseline
02Estimate automation
03Value saved capacity
04Review return
System disclosure · AI boundary · evidence
System typeDeterministic business-case model
AI boundaryGenerative model: none. The result comes from visible arithmetic applied only to the visitor’s assumptions.
Human gateThe business owner validates the time study, confirms which savings can actually be realized, and adds omitted implementation or risk costs before making an investment decision.
Current evidenceThe model covers positive, zero-cost, break-even, negative-return, invalid-input, and boundary scenarios. It makes every formula and excluded cost visible beside the result.
Interactive planning model · calculated locally
Editable assumptions · instant estimateAI automation ROI calculator
Your organization supplies the staffing, cost, time, automation, realization, and vendor-cost assumptions.
Hours, value, payback, and ROI are derived instantly from those inputs. They are not observed financial results.
Primary sources improve cost definitions and implementation controls; their benchmarks never overwrite your scenario.
Scenario readout
Positive modeled capacity
The estimated labor value recovered is greater than recurring software cost. Validate with a small pilot before purchasing.
Gross monthly labor value: $12,776Across the selected team
Realized labor value − software
Monthly net estimate × 12
Annual recurring value − setup
Setup cost ÷ monthly net savings
Time to recover one month of software cost
Year-one net value ÷ year-one investment
Recurring software + support ÷ modeled hours saved
Minimum automation needed to cover recurring cost
Current repetitive-work baseline585 hours / month65% recoverable · 80% value realization
Executive decision brief
Measure before purchase
Directional recommendation · recalculates with every scenario.
01Decision
Why this is the current recommendation
The model shows $11,576 in monthly net value, but its task volume, automation rate, and value realization are still planning assumptions.
Next best action
Time one representative workflow cycle and document task mix, exceptions, rework, and the destination of recovered capacity.
02Sensitivity
Transparent downside case
50% automation · 60% value realization · $6,171 monthly net value.
Boundary
Calculated scenario: 15 percentage points less automation and 20 points less value realization. It is a sensitivity test, not a forecast or probability interval.
03Pilot
Prove value before expansion
Choose one stable, high-volume workflow; preserve the current process as a manual fallback and exclude staffing changes from the pilot.
Measure
Observe at least one representative baseline cycle and one comparable pilot cycle using the same task mix. Log run time, human review, exceptions, rework, incidents, and every recurring operating cost.
Scale gate
Scale only when observed monthly net value stays positive, cycle time improves for comparable work, quality does not worsen, and no unresolved high-severity control issue remains.
Executive scorecard
Three KPIs that connect automation to business value
Definitions are operational; targets should be approved from the organization’s measured baseline.
01KPI
Observed monthly net value
Definition: Observed hours removed × approved total employer cost × verified redeployment share − software, support, monitoring, and administration cost.
Decision use
Primary economic KPI: confirms whether the automation creates usable capacity or avoided expense after all recurring costs.
Guardrail
Do not count idle capacity as cash savings or count the same recovered hour in more than one benefit category.
02KPI
Median end-to-end cycle time
Definition: Median completion timestamp minus intake timestamp for comparable cases, reported for baseline and pilot cohorts.
Decision use
Shows whether customers or staff actually receive a faster outcome, not merely fewer clicks inside one step.
Guardrail
Compare the same task mix and report the 90th percentile so difficult cases are not hidden by a better median.
03KPI
Exception and rework rate
Definition: Cases needing correction, override, escalation, or repeat processing ÷ completed pilot cases.
Decision use
Quality guardrail: detects whether apparent speed or savings shifted work into review, correction, or downstream risk.
Guardrail
Do not scale when severe incidents are unresolved or when rework rises beyond the pre-agreed tolerance, even if modeled ROI is positive.
Evidence boundary: Employees, cost, task time, automation, realization, and vendor cost are user inputs. Financial outputs and the downside case are deterministic estimates. Only a controlled before-and-after measurement can establish realized operational or financial impact.
Scenario-specific operating plan
Improve cost efficiency without overstating savings
Early estimate selected · recommendations recalculate with every input.
01Do first
Replace early estimates with an observed baseline
585 monthly repetitive hours are user-entered assumptions, not measured facts.
Next action
See the official basisSample task count, touch time, wait time, exceptions, and rework before committing to a vendor or changing staffing.
02Then
Give recovered capacity a named destination
20% of potential labor value—about $3,194 per month—is intentionally excluded from savings.
Next action
See the official basisAssign the recovered hours to a measurable queue such as avoided overtime, faster response, additional cases, quality review, or deferred hiring; do not call unused capacity cash savings.
03Protect
Set a recurring-cost guardrail
Software currently uses 9.4% of modeled realized value, leaving 91% before setup and any omitted operating costs.
Next action
See the official basisTrack licenses, usage, support, maintenance, and internal administration together; totals above $12,776 erase the modeled monthly net value.
04Protect
Keep quality, security, and human override in the ROI gate
The financial formulas do not measure incorrect outputs, privacy exposure, downtime, or the labor needed to review exceptions.
Next action
See the official basisDefine acceptable error and exception rates, access controls, approval points, monitoring, incident handling, and a manual fallback before scaling.
05Verify
Confirm the hourly input is total employer cost
$42 per hour is user-provided. BLS measures employer cost as wages plus benefits; its March 2026 private-industry average was $46.60 per hour, but national averages can differ sharply from this workflow.
Next action
See the official basisUse role-specific wages, employer-paid benefits, payroll taxes, and other included labor costs from finance or HR. Keep the BLS figure as context, not an automatic replacement.
Review formulas, assumptions, and limits Transparent model
- Baseline monthly hoursemployees × repetitive hours per employee per week × 52 ÷ 12
- Monthly hours savedbaseline monthly hours × estimated automation percentage
- Realized monthly labor valuemonthly hours saved × hourly labor cost × realized-value percentage
- Monthly and annual recurring net valuerealized labor value − recurring software cost; monthly net × 12
- Year-one ROI(annual gross labor value − annual software cost − one-time setup cost) ÷ (annual software cost + one-time setup cost) × 100
- Project break-evenone-time setup cost ÷ monthly net savings after recurring software; not reached when monthly net savings is zero or negative
- Monthly software recoverymonthly software cost ÷ gross monthly labor value × 30.4 days; shown when recovered value meets or exceeds the recurring monthly cost
This directional estimate assumes 52 working weeks spread across 12 months. “Realized value” discounts recovered time that cannot be redeployed into useful work. “Software cost per saved hour” excludes setup cost. The optional setup input can represent implementation, integration, and training, but the model still excludes maintenance outside the entered recurring cost, payroll burden not included in the hourly rate, taxes, downtime, error risk, ramp time, and changes in task volume. Add those costs to a full business case and validate both percentages with a measured pilot.
Primary-source registry
Reviewed July 15, 2026What each official source supports—and what it does not
Employer Costs for Employee Compensation — March 2026
Released June 12, 2026 · latest available on July 15, 2026Supports: Use total employer compensation—not wages alone—when estimating the hourly value of labor capacity.
Boundary: The $46.60 private-industry average is national context only; it does not replace an organization-specific cost.
Open official sourceRobotic Process Automation Playbook
Official guidance · accessed July 15, 2026Supports: Process selection, business-value measurement, license management, monitoring, error handling, and employee redeployment.
Boundary: Federal RPA guidance is used as an implementation reference, not as proof that a private workflow will achieve savings.
Open official sourceAI Risk Management Framework Playbook
Official page updated June 10, 2026Supports: Govern, map, measure, and manage AI risk; test, document, monitor, and retain appropriate human oversight.
Boundary: The voluntary framework improves decision discipline; it does not certify a tool or guarantee its financial return.
Open official sourceMethod, safeguards & evaluation
How the system works
A browser-local calculation converts weekly repetitive hours into a monthly baseline using 52 weeks divided by 12 months, applies the chosen automation percentage, values the saved capacity at the entered hourly labor cost, and separates recurring software cost from optional one-time setup cost before calculating payback and annual return.
Human control & safety
Saved hours are presented as capacity value—not guaranteed cash savings or headcount reduction. The optional setup input can represent implementation, integration, and training; maintenance, error, adoption, and change-management costs still require a full business case.
Evaluation plan
Formula tests · boundary validation · zero-cost handling · negative-return states
Live official research + fixturesIndependent exploration
/08CareSignal
A healthcare lead-scoring and ICP qualification system
CareSignal turns organization-level healthcare firmographics and buying-signal evidence into an inspectable 0–100 fit score, role-based buying-committee suggestions, a qualified-account dashboard, and a reviewable HubSpot handoff.
Which healthcare organizations deserve human account review, and why?
Share of surfaced qualified accounts that a human reviewer accepts for the next approved sales step.
Revenue teams can focus scarce research time on inspectable organization-level fit evidence instead of opaque scores or scraped contact lists.
01Identify account
02Enrich evidence
03Score ICP fit
04Approve CRM handoff
System disclosure · AI boundary · evidence
System typeSource-aware healthcare revenue-operations prototype
AI boundaryGenerative model: none. Official-registry retrieval, match ranking, weighted scoring, and role recommendations are deterministic and inspectable; unknown facts are not generated.
Human gateA sales or operations owner reviews the evidence, adjusts the account assumptions, and explicitly approves any production CRM handoff.
Current evidenceAutomated checks cover official-source merging, ambiguous matches, partial outages, five-digit ZIP-to-ZIP+4 matching, unknown employee totals, score totals, qualification thresholds, and CRM payload provenance.
Method, safeguards & evaluation
How the system works
A best-effort research step searches CMS NPPES, the CMS Provider Data Catalog, and GLEIF in parallel, ranks possible organization records, and requires a person to choose the right entity or location. Verified fields, user-entered estimates, unknowns, and synthetic fixtures remain distinct before deterministic scoring and a browser-local CRM preview.
Human control & safety
The workflow accepts organization-level data only—never patient data or PHI. Public registries do not provide a reliable enterprise employee total, so unsupported counts stay unknown; scores are prioritization aids, suggested decision-makers are roles rather than scraped contacts, and the HubSpot action remains a local preview.
Evaluation plan
Golden-account fixtures · score reproducibility · threshold behavior · missing-evidence warnings
Live portfolio simulationIndependent exploration
/09RoutePilot
An explainable AI inbound lead-routing workflow
RoutePilot transforms a structured inbound inquiry into transparent entered company context, an inquiry category, planning-value band, sales-owner assignment, CRM preview, personalized reply draft, team alert, and response-time trace.
Who should own an inbound inquiry, what response clock applies, and what information must be clarified before a reply?
Elapsed time from inquiry receipt to the first approved response from the correct owner with required context present.
Prospects reach the right human sooner while sales teams spend less time re-reading, forwarding, and reconstructing inquiry context.
01Capture inquiry
02Validate + classify
03Estimate + route
04Draft + alert
System disclosure · AI boundary · evidence
System typeBrowser-local inbound orchestration prototype
AI boundaryGenerative model: none. Classification, delivery-hour estimates, routing, and response templates are deterministic and browser-local; unknown domains are not researched or embellished.
Human gateThe assigned representative verifies the account, edits the response, and approves CRM and communication actions before anything leaves the workflow.
Current evidenceThe evaluation checks ordinary-domain fallback, competing-category scoring, every estimate driver, budget independence, routing precedence, regional ownership, SLA states, and the no-transmission boundary.
Method, safeguards & evaluation
How the system works
A browser-local orchestration model normalizes the company domain, scores competing inquiry signals, estimates implementation hours from rollout scope and operational complexity, and routes the opportunity by risk, complexity, category, and region. It produces reviewable CRM, response, and alert previews without transmitting the inquiry.
Human control & safety
No CRM record, email, or team alert is sent from the public prototype. Headcount is excluded from pricing, budgets never inflate the modeled range, unknown company evidence stays unknown, and every external action remains a reviewable preview.
Evaluation plan
Estimate monotonicity · budget independence · routing precedence · fallback ownership
Live public data + fixturesIndependent exploration
/10AccountBrief
A source-aware automated account-research assistant
AccountBrief converts a company name or domain into a structured research brief with an official-record summary, industry and available size context, pain-point hypotheses, buying-committee roles, buying-signal checks, discovery questions, need-aware drafts, and claim-level confidence warnings.
Is there enough verified account evidence to plan discovery, and which unknowns should a seller investigate next?
Elapsed time from an approved account-research request to a human-reviewed brief ready for its stated use.
Sellers spend less time assembling public records and more time validating relevant questions without turning hypotheses into company facts.
01Resolve company
02Assemble evidence
03Draft hypotheses
04Review provenance
System disclosure · AI boundary · evidence
System typeSource-aware public-company research assistant
AI boundaryGenerative model: none. Official records come from SEC EDGAR and exact GLEIF matches; first-party homepage fields stay labeled self-published; one deterministic evidence-aware planner creates questions and purpose-specific drafts.
Human gateThe seller opens and verifies the cited sources, corrects stale or missing evidence, and edits the discovery and outreach material before use.
Current evidenceAutomated checks cover organization and domain ranking, SEC and GLEIF enrichment, source links, withheld employee totals, ambiguity abstention, distinct need-aware drafts, and honest not-found and provider-error states.
Method, safeguards & evaluation
How the system works
A bounded server route uses Wikidata only for candidate resolution, checks SEC EDGAR and GLEIF for exact official records, and inspects bounded first-party homepage metadata when available. Synthetic fixtures remain available for reproducible evaluation; provider outages, unsupported employees, and buying signals stay visible as gaps, while hypotheses and drafts remain separate from sourced facts.
Human control & safety
Decision-makers remain role hypotheses, not scraped contacts. Public facts link to their sources, employee counts retain their point-in-time qualifier when available, missing buying signals remain missing, and outreach requires human review.
Evaluation plan
Entity disambiguation · legal-name collision rejection · provider-outage states · source coverage · company-specific plans
Seven of the ten products retrieve latest-available public data and show their method. HeatSignal uses Open-Meteo weather, the official IQAir AirVisual API for the nearest live city-level US AQI observation, NWS heat-index logic, AirNow color bands, and an OpenStreetMap view; Voltline combines EIA-930 and monthly state benchmarks with a versioned 2020 Census ZCTA, 2024 ACS occupied-housing, and final 2024 EIA-861 reference layer for clearly labeled ZIP-linked estimates and utility candidates; FieldLens uses NHTSA and routes title-history checks to DOJ-approved NMVTIS providers or NICB; SceneCraft inspects a submitted public page, authoritative RDAP registration data, Cloudflare malware-filter DNS, and Mozilla Observatory results; PlateScout uses Nominatim, Overpass, and OpenStreetMap place tags; AccountBrief uses Wikidata only to resolve an identity candidate, checks SEC EDGAR and exact GLEIF legal-name records, and labels bounded company-homepage fields as self-published; CareSignal uses CMS NPPES, the CMS Provider Data Catalog, and GLEIF for organization research while keeping synthetic fixtures visibly separate. PaperPatch, the Automation ROI Calculator, and RoutePilot process operational inputs inside the browser; the ROI calculator also links its planning guidance to BLS, GSA, and NIST. No revenue workflow sends data to a CRM, email, or team channel. Public data can be delayed, revised, incomplete, or unavailable, browser PDF support varies by document, and modeled workflow outcomes do not guarantee production performance or realized savings. Voltline’s ZIP-linked usage is not measured consumption, its county utility matches are not confirmed providers, and its annual averages are not tariffs or offers. These are independent products—not client engagements, emergency guidance, professional or financial advice, safety guarantees, popularity claims, identity verification, secure redaction, or production performance claims.
How I work
Design the loop,
not just the prompt.
The strongest AI experiences make context, state, uncertainty, permissions, and human judgment visible—not magical.
Frame
Define the user decision, source of truth, acceptable failure, and what should remain human.
Ground
Connect trusted context, tools, and permissions. Make citations and provenance part of the product.
Prototype
Build the smallest end-to-end workflow and make the system’s state visible in the interface.
Evaluate
Test quality, cost, latency, safety, and failure modes against representative tasks and a non-AI baseline.
Improve
Add feedback loops, human escalation, monitoring, and versioned changes before expanding autonomy.
Field notes
What the builds
made visible.
Short, practical notes about designing evidence, uncertainty, privacy, and human control into working systems.
/01
Designing an honest partial audit
What SceneCraft shows when public page evidence is blocked, incomplete, or unsafe to retrieve—and why unknown must stay unknown.
Read field note ↗/02Fresh data needs visible time
How Voltline and HeatSignal separate fetch time, observation time, source delay, and the decision a visitor still owns.
Read field note ↗/03Local-first PDF work has boundaries
Why keeping a file in the browser improves privacy without turning a visual cover-up into secure redaction or identity proof.
Read field note ↗Start with the problem
Building
with AI?
If you’re exploring a knowledge product, workflow automation or agentic design, a multimodal experience, or a role where this thinking matters, I’d like to hear what needs to work—and what cannot fail.
01 Hiring and roles02 Product collaboration03 General networking
Available for focused product design, rapid prototyping, and AI system collaboration from Texas, USA. Serious inquiries receive a personal reply.
Project inquiry / Google Forms handoff