Independent product case study / Healthcare information and navigation
Care Navigator AI
Care Navigator AI is a source-backed U.S. health-information product for understanding medical terms, reviewing patient documents, checking official medicine-label evidence, organizing questions, and comparing carefully labeled cost benchmarks. It is designed for preparation and navigation—not diagnosis, prescribing, emergency care, or a promise of coverage or price.
01 / Brief
A person may have a discharge document, an unfamiliar diagnosis, a medication name, and a cost question at the same time. Search results often mix source quality and omit the next practical question. The product must organize evidence without turning general information into individualized medical advice.
The interface prioritizes plain language, large controls, keyboard and screen-reader access, visible source dates, emergency escalation, and printable preparation. It assumes limited technical confidence and makes uncertainty explicit.
02 / Architecture
Each stage has a bounded responsibility. The interface preserves source context and leaves the final judgment with the person using it.
The user chooses health search, document review, costs and coverage, or preparation instead of entering an unbounded clinical prompt.
Task-specific inputs · no account required · no clinical diagnosis outputServer routes query official or trusted sources close to each claim, such as NLM, NIH, FDA, CMS, and RxNorm.
Bounded requests · schema validation · explicit source failuresPDFs and images are converted into reviewable text and candidate medical terms; low-confidence medication identity requires user confirmation.
OCR and text extraction · privacy notice · no silent medication guessCondition summaries, medicine labels, coverage directories, and price benchmarks retain different labels and limitations.
Official label ≠ prescription direction · benchmark ≠ pharmacy quoteResults become questions, source links, warning signs, and a printable plan rather than a treatment order.
Plain language · expandable detail · large actionable controlsEmergency symptoms route to 911 or 988 where applicable; medication instructions route to a pharmacist or prescriber; coverage and final price route to insurer and pharmacy.
Human escalation is a product outcome, not a failure03 / Product decisions
Tradeoffs made visible
The product may return a narrower answer or an explicit no-match rather than synthesize unsupported health claims.
The first view is concise and plain-language; full source detail, administration sections, and provenance remain available on demand.
NADAC and CMS data can show product-level benchmarks, but the interface does not claim a live store quote or plan-specific copay.
A slightly slower identity step is preferable to showing label evidence for the wrong medicine, strength, or dosage form.
Failure is a designed state
The review surfaces extraction confidence and asks for a clearer image or manual confirmation. It does not silently invent missing directions.
The search broadens across trusted rare-disease and biomedical indexes, ranks exact condition records first, and labels missing public definitions rather than creating one.
Partial and mismatch states are visibly separated; mismatches stay hidden by default and the user is directed to confirm the final price with a pharmacy and insurer.
Useful verified sections survive where possible. Unavailable fields show a source-specific recovery path and never become generated facts.
04 / Evaluation
This is the evaluation surface documented by the implementation—not a claim that every target has already been met.
Exact diseases, medications, tests, rare conditions, and common misspellings should rank the correct clinical concept above administrative records.
Exercise selectable and scanned PDFs, JPG, PNG, screenshots, rotation, low contrast, multi-page files, and deliberately unreadable inputs.
Verify identity confirmation, label dates, boxed warnings, interaction framing, pharmacist escalation, and refusal to turn general label text into a personal dose.
Test exact, partial, and mismatch combinations for name, RxNorm concept, strength, dosage form, benchmark date, and missing plan-specific coverage.
Review keyboard flow, zoom, landmarks, labels, focus visibility, reduced motion, error announcements, contrast, and plain-language recovery on mobile and desktop.
Adversarial checks must confirm the product does not diagnose, prescribe, promise a cure, guarantee a price, or delay emergency care.
05 / Build ownership
AI accelerated drafting, implementation alternatives, and adversarial review. Product judgment, source selection, interface decisions, testing, and release approval remained mine.
Defined the non-diagnostic safety boundary and the decisions that remain with clinicians, pharmacists, insurers, and emergency services. Designed the information architecture across health search, document review, medicine evidence, costs, providers, and preparation planning.
Exploring implementation options, generating draft test cases, checking edge cases, and accelerating repetitive code—always reviewed against the product boundary.
Retrieval precision, Document extraction. Failures were converted into visible recovery states and narrower claims.
Trust improves when missing evidence, partial provider failures, and the human decision are designed as first-class states instead of footnotes.
Clarify the task → Retrieve authoritative evidence → Extract with confirmation → Separate evidence types → Organize the next conversation → Escalate the right owner.
The review surfaces extraction confidence and asks for a clearer image or manual confirmation. It does not silently invent missing directions.
06 / Provenance
Condition, test, and medicine education with U.S. National Library of Medicine provenance.
Patient-oriented rare-disease information and links to current research and support resources.
Official product-label sections, dates, warnings, adverse reactions, and administration text; not a substitute for a prescription label.
Normalized medication concepts used to reduce ambiguity across generic name, strength, and dosage form.
Public reimbursement and acquisition-cost evidence used only with dataset-level limitations and match labels.
Every surface documents what its source can establish, what it cannot establish, freshness behavior, and the human verification owner.
Implementation note
Reviewed against the production source contracts, extraction routes, medication evidence, cost matching, accessibility states, and automated release suite. This is product verification, not clinical validation or a substitute for professional medical review.
See the system under real conditions