Portfolio project / Evidence-first sales research
AccountBriefAutomated account-research assistant.
Turn a real company name or domain into a sourced, reviewable account brief—without turning missing data into confident fiction.
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.
01 / Interactive system
Research with a visible evidence boundary.
Real-company searches use bounded identity discovery, SEC EDGAR, GLEIF, and a company’s public homepage when available. Results distinguish official records, self-published first-party context, discovery-only identity data, user goals, hypotheses, provider outages, and missing fields. No personal contacts or buying signals are invented.
Loaded the offline synthetic Northstar Health Systems fixture.
Northstar Health Systems
northstar-health.example
Multi-site outpatient care network
This fictional fixture represents a regional healthcare operator coordinating patient access, referral intake, and revenue-cycle work across a growing clinic footprint.
Healthcare operations
The fixture consistently describes outpatient services, patient-access workflows, and revenue-cycle operations.
500–1,000 employees
A fictional careers snapshot and clinic directory support a range rather than an exact headcount.
Possible pain points
Referral intake may require duplicate entry
The fixture’s operations note describes fax, portal, and phone referrals converging into one central team, suggesting avoidable re-keying and status follow-up.
Eligibility checks may constrain scheduling
A synthetic job description emphasizes manual benefit verification before appointments and escalation of incomplete records.
Decision-maker role hypotheses
VP, Patient Access
Likely workflow owner for referral intake, scheduling quality, and access-center performance.
Chief Information Officer
Likely technical stakeholder for integration, security review, identity, and clinical-system constraints.
Buying signals, with evidence attached
Patient-access automation role opened
The synthetic careers snapshot includes a newly listed director role responsible for workflow standardization and automation pilots.
Two clinics added to the network
The fictional organization update describes expansion that could increase intake volume and process variance.
Questions before conclusions
These are questions—not assertions. They turn hypotheses into a measurable discovery conversation.
- 01
How many referrals arrive through each channel, and where is the same information entered more than once?
- 02
Which eligibility exceptions consume the most staff time before an appointment can be confirmed?
- 03
What accuracy, privacy, and human-review thresholds would an automation pilot have to meet?
- 04
How would recovered capacity be measured without assuming it becomes immediate payroll savings?
Use for product demonstration only
The fixture shows the full workflow with 100% source-indexed synthetic claims, but it does not describe a real company or current commercial opportunity.
Next best action: Use the fixture to evaluate workflow behavior, then repeat the process on a correctly resolved public organization before making a real business decision.
Evidence boundary: Wikidata may assist candidate discovery; SEC EDGAR and exact qualified-name GLEIF matches support selected official facts; bounded homepage fields remain self-published first-party context. None of these sources proves a current initiative, pain point, budget, decision-maker, or buying signal. Role suggestions, business problems, and outreach language remain hypotheses until a human verifies them; nothing is sent or saved automatically.
- 01
100% current factual-claim coverage
3 of 3 displayed factual claims point to a loaded source ID. Coverage shows whether displayed factual claims have source IDs in this brief. It does not prove that the sources are complete, current, or sufficient for a commercial conclusion. - 02
Shadow-pilot mode
Run AccountBrief beside the current research process on a consecutive, representative account cohort. Keep existing account decisions and outreach controls unchanged while reviewers record match errors, unsupported claims, corrections, preparation time, and whether the brief was usable. - 03
Measurement
Compare source-backed coverage, reviewer corrections, and time to an approved internal brief by company type, source coverage, and match-confidence band. Track not-found and abstention cases as intentional safety behavior. - 04
Expansion gate
Choose targets from the organization’s measured baseline. Expand only when reviewers approve the evidence quality and time savings without a rise in identity errors, unsupported claims, privacy issues, or unreviewed outreach.
Measure usefulness without rewarding unsupported claims.
Source-backed factual claim coverage
Definition: Displayed factual company claims with at least one traceable approved source ID ÷ all displayed factual company claims.
Material reviewer correction rate
Definition: Briefs requiring removal or correction of an identity, factual, timing, recipient, or evidence claim ÷ all briefs reviewed.
Median time to approved internal brief
Definition: Median elapsed work time from research start to human approval of an internal evidence brief, with the 90th percentile also reported.
Prospecting email, with a human gate.
- Confirm the organization match and every source-backed fact.
- Replace the role archetype with a verified recipient only after human review.
- Remove any premise the source registry does not support.
- Keep business outcomes conditional until a baseline is measured.
Every sourced fact points back here.
Synthetic service-line profile
Fictional description of services and operating model.
Synthetic clinic-network update
Fictional footprint and expansion update.
Synthetic careers snapshot
Fictional roles and responsibilities; not a live job board.
Synthetic operations interview
Fictional workflow notes created for evaluation.
Loaded source IDs: NS-01 · NS-02 · NS-03 · NS-04. Fixture content is fictional and exists only to demonstrate product behavior.
Next safe step: after a human verifies the organization and inquiry context, open RoutePilot to prepare a browser-local routing decision and review-only handoff.
02 / Accessible architecture
A pipeline built to stop unsupported claims.
Read left to right on larger screens or top to bottom on mobile. The ordered list remains understandable without its visual connectors.
- 01
Input boundary
Accept a company name or public domain; normalize it, reject personal emails and private hosts, and cap request size.
- 02
Entity resolution
Prefer exact root-domain evidence, reject same-name legal-entity collisions, and cross-check SEC aliases or exact qualified GLEIF names.
- 03
Evidence registry
Preserve regulator, legal-entity, and first-party links with retrieval time; leave missing or unverified fields unknown.
- 04
Need-aware drafting
Use one evidence-aware planner to turn company context, hypotheses, gaps, and the user’s goal into the visible questions and selected deliverable.
- 05
Claim guard
Expose field-level confidence, warnings, point-in-time qualifiers, limitations, and an honest not-found path.
- 06
Human review
A person verifies the company, evidence, recipient, tone, and proposed next step before any outreach.
03 / Product walkthrough
See the research boundary in motion.
The recorded walkthrough demonstrates the offline fixture path, claim-level evidence, outreach review, and safe abstention. The live interface above now adds sourced public-company resolution.
Read the written video summary
The recording opens an offline synthetic account fixture, reviews its summary and source IDs, checks pain-point and decision-role hypotheses, reads buying-signal warnings, copies a draft for human review, then searches an unmatched company to show that unsupported claims are withheld. Live public-company lookup was added after this recording and is available in the interactive section above.
04 / Fixture evaluation
Measure the behavior that earns trust.
Each fixture resolves through its exact company name and canonical domain.
Every offline-fixture claim carries confidence, at least one source ID, and a verification warning.
No fixture or fallback contains an invented person, email address, or phone number.
Each fallback withholds industry, size, pain points, roles, signals, and personalized outreach.
The fixture suite verifies deterministic behavior, while mocked route checks cover entity disambiguation, same-name legal-entity rejection, SEC aliases, root-domain preference, bounded first-party structured data, provider-state separation, company-specific plans, and safe abstention. It does not prove that public records are complete or current, nor does it prove sales conversion or production readiness.
05 / Technology
Small architecture.
Strong boundaries.
- Next.js App Router
- React + TypeScript
- CSS Modules
- Deterministic fixture resolver
- SEC EDGAR data APIs
- GLEIF Legal Entity API
- First-party Organization JSON-LD
- Evidence-tier + provider-state ledger
- Bounded identity resolution
- Organization entity scoring
- Source-indexed claim schema
- Browser Clipboard API
- Node test runner
- Cloudflare-compatible server route
06 / Safeguards
What the system
refuses to do.
No fabricated contacts
The prototype never invents names, email addresses, phone numbers, or reporting relationships.
Bounded public website inspection
An exact public domain may supply homepage metadata or Organization JSON-LD. Fetches stay HTTPS, follow a short same-registrable-domain redirect chain, enforce time and size limits, and run behind a public-network-only fetch policy.
Unknown means unknown
An ambiguous or unmatched organization receives no invented facts, buying signals, people, or personalized outreach.
Claims stay inspectable
Public-source facts include confidence, source IDs, retrieval time, and a warning about what still needs verification.
Hypotheses look like hypotheses
Possible workflows and role archetypes are visibly separated from sourced company facts.
Human approval is final
Nothing is sent, saved to an account, or acted on automatically.
Project handoff