Projects
Projects
Applied-AI systems, healthcare-AI first, each with the technical decisions behind it and a results table where every metric shows its qualifier, most with an architecture diagram. The failures are in here too, on purpose.
Clinical Ops Copilot
An agentic system that reads a patient's chart, checks it against payer rules, and prepares prior-authorization paperwork for a human to approve.
- Decision macro-F1
- 93.7%Decision macro-F1n=16, locked held-out split, synthetic data
- X12 278 round-trip decision agreement
- 16/16X12 278 round-trip decision agreementself-consistency test on the locked split; measures ingestion fidelity, not clinical accuracy
- healthcare
- agents
- evals
- infra
ChartExtractor
Turns a doctor's free-text oncology note into a structured record, and tells you exactly where it isn't sure.
- Macro-F1, CI gold set
- 93.6%Macro-F1, CI gold set6-note synthetic CI gold set, deploy gate at 85%
- Macro-F1, real notes
- 53.8%Macro-F1, real notes50 hand-labeled MTSamples notes; ~40-point gap vs synthetic, disclosed publicly
- healthcare
- evals
- infra
handoff-lens
A design-to-code readiness assistant that reads a Figma file through the public Figma REST API, runs deterministic readiness checks, and writes a customer-facing summary.
- Precision / recall, deterministic checks
- 1.000 / 1.000Precision / recall, deterministic checkson its 6-fixture self-authored set; 6/6 fixtures exact, 11/11 planted findings, 0 false positives
- All 6 LLM summaries, live run
- schema-validAll 6 LLM summaries, live runliveness and schema-validity only, explicitly not an accuracy score
- design-tools
- agents
- evals
LedgerSim
A weekend-scale core-banking ledger with a double-entry engine, product-contract hooks, and a Claude layer that turns plain-English product specs into validated parameters.
- Field checks correct
- 114/114Field checks correcton its 19-spec self-authored golden set, single live run, zero misses
- Surfaced live, each regression-tested
- 3 bugsSurfaced live, each regression-testedfound by running the eval live rather than assuming it passed
- fintech
- agents
- evals
RepSignal
A TypeScript and Node.js conversation-intelligence integration: a webhook ingests a sales-call transcript, a shared Zod schema validates it, and Claude under forced tool-use returns a schema-validated coaching scorecard.
- LLM-judged field checks
- 66/70LLM-judged field checkson its self-authored synthetic 14-transcript eval set, one live run; representative not fixed
- Talk/listen ratio, deterministic
- 14/14Talk/listen ratio, deterministiccomputed in code, exact by construction; reported separately from the 70 LLM-judged checks
- typescript
- integrations
- agents
- evals
Trust Marketplace
A Medicaid referral platform connecting healthcare providers and community-based organizations, built from the first customer conversation to a live platform.
- Outreach messages processed
- 5.12M+Outreach messages processedplatform total, Trust.care reporting
- Cohort engagement
- 19.2%Cohort engagement~12.8x the published CHCS 1.5% Medicaid SMS benchmark, in a 63.8%-no-email cohort
- healthcare
- agents
ChartExtract-UI
A single-screen human-in-the-loop console for reviewing, correcting, and approving what an extraction model pulled from a clinical note.
- Tests passing
- 55Tests passingVitest + React Testing Library, CI-gated on every push
- Review-threshold defaults, live vs demo mode
- 0.75 / 0.90Review-threshold defaults, live vs demo modelive matches the API's server default; demo raised because fixture confidences cluster around 0.85
- frontend
- healthcare
AWS Claims Pipeline
A serverless claims-intake pipeline defined entirely in Terraform, API Gateway to Lambda to SQS to DynamoDB, deployed for real to AWS with least-privilege IAM, exercised end to end, and then destroyed with the evidence captured.
- deployed for real, then destroyed
- 27 resourcesdeployed for real, then destroyedus-east-1; apply reported 27 added, destroy reported 27 destroyed, then CLI-verified zero residue
- total cost incurred
- under $0.01total cost incurredserverless free-tier only, DynamoDB on-demand, roughly 30 Lambda invocations, captured in EVIDENCE.md
- infra
- AWS
- Terraform
DPO Preference Sketch
A pre-registered offline DPO study that teaches a small model to abstain on unanswerable clinical questions instead of fabricating, with the DPO loss hand-implemented because the library trainers crashed on Apple Silicon.
- Held-out abstention recall
- 0.167 to 1.000Held-out abstention recallof 24 unanswerable items, 4/24 to 24/24; agreement with self-authored labels on a frozen 60-item synthetic set, not a general accuracy claim
- Held-out behavior accuracy
- 0.667 to 0.983Held-out behavior accuracyabstain iff unanswerable, +0.317, single run on a self-authored synthetic set, met the pre-registered success bar
- evals
- healthcare
- RLHF
EHR Phenotype Sketch
A PheKB-style computable Type 2 Diabetes and Hypertension phenotyping pipeline over synthetic Synthea EHR data, combining ICD-10-CM, SNOMED, LOINC, and RxNorm value sets with explicit exclusion logic and a two-implementation concordance check.
- Type 2 Diabetes cases (12.0%)
- 13 / 108Type 2 Diabetes cases (12.0%)synthetic Synthea cohort; an order-of-magnitude sanity check against ~9 to 11.6% US prevalence, not an accuracy claim
- Two-implementation concordance
- 20 / 20Two-implementation concordancestratified N=20 (all 13 cases plus the first 7 non-cases by id); two implementations of the same spec agreeing, not clinician chart review
- healthcare
- clinical informatics
- SQL
clinical-rag
write-up comingA hybrid-retrieval clinical RAG system with reranking, a RAGAS evaluation harness, and CI. No accuracy metrics claimed yet.
- healthcare
- evals
- infra
Sketches
Sketches
Smaller, focused builds. Each one closes a single specific gap end to end, on synthetic or self-authored data, and the depth lives in the repo rather than in a case study here.
- snowflake-claims-sketchA Snowflake-dialect claims warehouse over 240 synthetic claims: VARIANT landing, QUALIFY dedup, MERGE upserts, and a Streams and Tasks DAG, with the denial-rate mart verified on a live trial account.
- recruiting-skill-sketchAn Agent Skill in Anthropic's open SKILL.md format, plus a reference implementation that turns screening notes into evidence-quoted scorecards. Over three self-authored synthetic notes, version 1 produced 3 of 3 grounded scorecards with no invented quotes and 12 of 15 exact score matches. The measured version 2 did not cure the rule miss and introduced a protected-attribute regression, scoring a candidate down on caregiving status where version 1 had abstained. Version 1 remains the version to use.
- dpo-preference-sketchA pre-registered offline DPO study that taught Qwen2.5-0.5B-Instruct to abstain on unanswerable clinical questions instead of fabricating, on 299 self-authored preference pairs. Measured as agreement with my own labels on a held-out 60-item synthetic set, abstention recall moved 0.167 to 1.000 and behavior accuracy 0.667 to 0.983. The DPO loss is implemented by hand from the objective.
- edi-835-goA Go port of my Python X12 835 remittance parser, with a cross-language parity harness that proves exact output equality over shared fixtures.
- ledger-stream-sketchA double-entry ledger on PostgreSQL and Redis Streams, with idempotent at-least-once consumers and a self-authored synthetic event-sequence eval set.
- ops-sql-forensicsA synthetic SaaS ops dataset with four planted incidents and 12 diagnostic SQL queries, each proven by a test that it finds its incident.
- secure-transfer-sketchA defense-in-depth clinical file-transfer demo: key-auth SFTP, PGP, and signed SHA-256 manifests, with verify-before-decrypt and tamper quarantine, on synthetic data only.
- integration-sync-sketchAn idempotent integration-sync engine reconciling calendar, email, and CRM-shaped stores into one timeline, with incremental cursors, a conflict audit trail, and a dead-letter queue.
- prompt-ab-harnessA pre-registered A/B test of two prompt strategies, with McNemar exact and bootstrap CI, reporting an honest null result.
- eval-reliability-sketchA small self-authored study of two evaluator-reliability failure modes: judge-agreement decay under distribution shift, and abstention versus hallucination.
Business case
Business case
One artifact in a different register: the case for a deployment rather than the deployment itself. The health system in it is fictional, and the artifact publishes the downside case, the missed gates, and the cash-share analysis that makes the central case cash-negative, next to the headline numbers.