AIE Discipline 01 / 04

AI Engineering

LLM systems that hold up in a regulated industry. Structured outputs, evidence rules, and humans where it matters.

The story

Most of my AI work has shipped inside Hybridge, a dental implant company. That means HIPAA, real patients, and zero tolerance for confident nonsense. The constraint shaped how I build: the model never free-writes. It fills a validated JSON schema, cites a verbatim transcript quote for every claim it makes, and gets one retry with a hardened prompt before the pipeline escalates to a human.

The pattern repeats across systems. The consultation grader scores doctors against their own 30-day average, not an arbitrary bar. The call-coaching platform gates every real-PHI path behind one compliance switch and keeps the patient's name and transcript out of every log line, so a demo or a test can't touch real data by accident. Patient de-duplication proposes matches with confidence scores but never merges on its own.

The interesting work isn't the prompt. It's everything around it: schemas, retries, idempotency, audit trails, and deciding where a human stays in the loop.

What I bring

Structured output, enforced

Response schemas with validation and retry loops. The model fills a contract; it doesn't write prose that gets parsed by hope.

Evidence-grounded scoring

Every score cites a verbatim quote and timestamp. If the model can't point to it in the transcript, it doesn't get to say it.

PHI-safe by architecture

De-identification before inference, BAA-covered model endpoints, no PHI in logs, audit trails in BigQuery. Compliance is in the data flow, not a policy doc.

Humans where it matters

Confidence thresholds route uncertain calls to review queues. AI proposes; people approve the irreversible stuff.

The projects

Built, shipped, and used.

Enterprise Search (RAG)

In build

An internal AI search over all of Hybridge's knowledge. Hybrid retrieval, reranking, deterministic conflict resolution, and citations, with an honest 'I don't know' when the evidence is thin. Built trustworthy-first and shipped stage by stage.

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
An eval harness (recall, nDCG, MRR, faithfulness) decides when complexity earns its place

Read the full story

Doc Coach: Consultation QA

In production

Every Zoom consult scored against a 7-criterion clinical rubric by Gemini with structured output, delivered as color-coded coaching reports to doctors and the CEO within minutes of the call.

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
Weekly and monthly auto-generated comparative insight reports across doctors

Read the full story

NPC Coach: Call Coaching

In production

A pipeline that coaches new-patient phone calls: it finds the relevant calls among thousands, scores each against the practice playbook with a verbatim quote behind every criterion, and writes per-coordinator trend reports. A full rebuild of an n8n prototype.

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
Booking rate is deliberately not a success metric. Correctness over conversion

Read the full story

CDF: AI-assisted diagnosis

In build

The diagnostic layer where AI reads the patient before the doctor does: it analyses CBCT, radiographs, intraoral scans, and clinical photos to surface decay, missing teeth, restorations, and bone-support patterns, then routes every finding to a doctor validation queue. AI assists, the doctor signs.

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
Standardized bone-support visualization: current architecture against an ideal reference

Read the full story

CRM intelligence layer

Design complete

Inside the CRM blueprint: LLM-proposed patient de-duplication with mandatory human review, and structured extraction of five thousand free-text activity logs with the originals preserved verbatim.

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