Projects.

Problem

Company knowledge is scattered across drives, decks, and people's heads, and in a HIPAA shop a confident wrong answer is a liability, not a demo bug. A basic chunk-embed-and-pray RAG fails on exact terms, on conflicting documents, and on access control.

Solution

A trust-first RAG. The read path filters by permission, retrieves with BM25 and pgvector together, fuses with RRF, reranks a bounded top-k, resolves conflicts deterministically, and answers with citations or abstains. The governed write path dedups, parses with layout awareness, chunks semantically, and stamps every chunk with its model and strategy versions.

Process

Built like a curriculum, one stage at a time, with an eval harness as the referee: recall, nDCG, MRR, and faithfulness measured against a managed baseline before any answer is trusted. The agentic corrective loop is gated behind a confidence check, so it never runs on a simple query.

Result

Foundation and auth shipped: domain-locked Identity Platform sign-in, roles and groups, Terraform, and tested CI/CD, all GCP-native and HIPAA-eligible. Ingestion is fully specified and in build; query serving, evaluation, and the front ends are designed and queued.

Problem

Bad CBCT scans (poor field-of-view, motion, missing anatomy) kept reaching the implant-design queue. A vendor wanted $98K up front + $26K/yr to fix it.

Solution

An in-house classifier. Frozen DentalSegmentator nnU-Net v2 encoder + a compact 500K-param head, packaged as a Cloud Run service that scores every scan in seconds.

Process

Bake-off across six architectures, then iterated through Models G, H, J as data taught more. A strict 20-scan CICT holdout caught a leaky AUROC of 0.80 in CI. Shipped the honest 0.6309 rather than game the gate. Tag-based CI/CD lets us hot-swap the head checkpoint via GCS with no OpenVINO IR rebuild.

Result

Under $50/mo running cost. ~5.5s per scan on CPU, scale-to-zero. The 10-year recurring loss from bad scans closes.

Problem

An earlier n8n grader proved the idea, scoring each call out of 60 from recordings it pulled off SFTP every hour. But it emitted a brittle semicolon-delimited blob into a single Google Doc, carried no transcript evidence, computed no real score, shared the patient's name and full transcript company-wide, and silently skipped any call when the workflow died mid-run.

Solution

A full rebuild on a HIPAA-scoped Google Cloud footprint. A pipeline ingests recordings from SFTP and CallRail, transcribes with Vertex Gemini, classifies the call, scores relevant new-patient calls on six criteria with a verbatim quote behind each, resolves a stable patient ID, and writes a validated verdict to BigQuery with the report in Cloud Storage.

Process

Ports-and-adapters so the same code runs on in-memory fakes in tests and real cloud in prod, gated by one NPC_BAA_ACCEPTED compliance switch. The scoring math is pure Python, not the model: the 0-100 headline is a 40/40/20 weighted sum, while three hard gates and a patient-urgency override ride alongside as loud flags, never a hidden score cap.

Result

First full day on live calls: ~285 recordings ingested, 40 real new-patient calls scored with quoted evidence, 3 voicemails auto-filtered, 28 flagged for missing a non-negotiable. 1,156 tests green. Live on Cloud Run behind Google sign-in, locked to the practice domain.

Problem

The old dashboard hit the Monday.com API live, broke constantly, and patient↔lead linkage sat at 49% / 65% across the two clinics. Leadership argued about whose numbers were right.

Solution

Apps Script on weekly Excel exports. One shared metrics module so every tab and widget computes from the same definitions. Six tabs share one filter bar.

Process

Found that Monday's connect column (a real board_relation, not a mirror) survives Excel export. Encoded every business rule explicitly: marketing taxonomy, location filter, excluded reasons, scheduled-date semantics, consult-show parsing.

Result

Linkage 49.5% / 65.3% → 98.8% / 99.9% across two boards. Weekly recon: half a day → 3 minutes. Surfaced orphan re-treatment patients worth ~$460k of treatment value.

Problem

Per-seat SaaS automations were expensive and getting more so. C-suite analytics relied on manual exports each week.

Solution

Self-hosted n8n on AWS via Docker. One platform for every team automation, with data flows standardized across the org.

Process

Migrated existing automations into n8n. Built executive analytics pipelines that publish on a schedule instead of on-demand.

Result

20% lower automation cost. 500+ hours/year recovered on executive analytics. AI initiatives ship on a shared substrate.