Work

AI Engineering. LLM systems that hold up in a regulated industry.

I build healthcare LLM pipelines that fill a schema, cite their evidence, strip patient data first, and send anything uncertain to a person.

AIE 7 projects · 5 articles

How I work.

Built for healthcare. My AI work ships inside a dental implant company: HIPAA, real patients, no room for confident nonsense.
The model fills a schema. It fills a validated JSON schema and cites a verbatim transcript quote for every claim.
One retry, then a person. A failed answer gets one retry with a hardened prompt, then the pipeline hands it to a human.
Patient data stays out. One compliance switch gates every real patient data path, and names and transcripts never reach a log.
Matches proposed, never merged. Duplicate patient matching suggests pairs with confidence scores and never merges on its own.
The prompt is the small part. The real work is schemas, retries, idempotency, audit trails and where a human stays in the loop.

Structured output, enforced

Schemas with validation and retries. The model fills a contract, not free prose.
Evidence-grounded scoring

Every score cites a verbatim quote and timestamp, or it doesn't get said.
PHI-safe by architecture

Data stripped before inference, BAA covered endpoints, no PHI in logs, BigQuery audit trails.
Humans where it matters

Uncertain calls go to review queues. AI proposes; people approve anything irreversible.

In build

An internal AI search over Hybridge's knowledge, with citations and an honest 'I don't know' when evidence is thin.

Hybrid retrieval (BM25 + pgvector) fused with RRF and reranked, not vector-only
Conflict resolution by rule: validity, authority, recency, else declare the disagreement
Group access enforced as a pre-retrieval filter, so restricted docs never leak into a search

In production

Every Zoom consult scored by Gemini on seven clinical criteria, with coaching reports to doctors within minutes.

Three-tier identity resolution: Zoom participants → transcript speakers → LLM extraction
Verbatim-quote requirement cut hallucinated feedback by ~35%
Idempotent at three layers: webhook, ledger append, Drive upload

In production

Finds the relevant new patient calls among thousands and scores each against the playbook, a quote behind every criterion.

A verbatim transcript quote behind every one of six scoring criteria
Hard gates are flags, not caps: a call can score 84 and still carry a red 'missed a non-negotiable' badge
Patient-urgency override: schedule a consult-only for someone in real pain and it's recorded as a safety failure

In build

AI reads CBCT, radiographs, scans and photos for decay, missing teeth and bone support; a doctor validates every finding.

AI-assisted findings on 2D radiographs and CBCT: decay, missing teeth, restorations, bone levels
Every AI finding stays provisional until a doctor confirms it in a validation queue
Signed corrections feed a continuous-learning loop on the practice's own validated cases

In production

A BigQuery warehouse joining six marketing sources, ad spend to treatment value, documented so LLMs read a guide first.

Every column described, plus a data guide written for LLMs to read first
A weekly Gemini eval on 22 golden questions with an 80% gate; the first run scored 77%, short of it
98.5% of patients linked to their lead; 99.5% join agreement with the weekly exports

Design complete

Duplicate patients proposed by an LLM with mandatory human review, plus structured extraction of five thousand activity logs.

No autonomous merges, ever. The data has 74 shared phone numbers (married couples)
Every parse carries a confidence score and source reference for re-runs
Consent tokens extracted into an append-only ledger