AI Engineer
Hybridge Implants, New York
- Consultation QA Pipeline. Cloud Run + FastAPI grades every Zoom consult against a 7-criterion rubric. Schema-validated Gemini scoring, three-layer doctor ID, HIPAA-clean without Workspace DWD. -35% hallucinations vs the no-schema baseline; +130% acceptance and +43% revenue downstream. Now also drafts the doctor's clinical note from the transcript, iterated with the clinical team and guarded so a truncated note can never ship.
- Pipeline observability hardening. Hardened that same QA pipeline after it kept showing ✅ complete while emails never sent and Sheet rows never appeared. Killed 9 silent-failure modes (bare-except graveyard, connection-level retry gaps, status-color drift), added durable webhook rows, a rerun-replaces-old flow, a consistency reconciler, and a plain-English Logs UI. 51/0/0 in the audit, +50 targeted tests.
- NPC Coach. Rebuilt a brittle n8n call grader into a coaching platform for the front desk: it finds every new-patient call, grades it against the practice playbook with a verbatim transcript quote behind each of six criteria, and tracks each coordinator's trend in BigQuery behind a 26-endpoint React dashboard. 40/40/20 weighted score plus 3 hard gates as flags, not caps; ports-and-adapters gated by one NPC_BAA_ACCEPTED switch so 1,156 tests run with no cloud and no PHI. Day one on live calls: ~285 ingested, 40 scored, 3 voicemails auto-filtered, 28 flagged.
- CBCT Scan Validator. In-house dental CT classifier. Replaced a $98K + $26K/yr vendor quote with a Cloud Run service under $50/mo. Frozen DentalSegmentator + multi-scale head, OpenVINO on CPU, ~5.5s end-to-end. 20-scan CICT gate on every push.
- Treatment Estimator. Next.js + Postgres rebuild of a decade-old quoting tool. Every saved line stores its captured price and price book version, so the 6-month price guarantee holds. Shipped end-to-end in ~1 month; a prior vendor never shipped in a decade.
- Cowork Dashboard. Weekly Monday.com exports turned into an always in sync dashboard, moved from Apps Script to Cloud Run in September with one definition per metric. Surfaced ~$460k of orphan patient value. Weekly recon: half a day to 3 minutes.
- Accounting Automation Suite. A dozen Python scripts that replaced the controller's manual weekly imports across Denticon, MagicTouch, Paychex, and two banks. About 6 to 8 hours a week down to 30 to 45 minutes, roughly 400 hours a year back.
- Reconciliation. A deterministic engine that cross checks the schedule, production ledger, and clinical notes across Eaglesoft and Denticon to confirm every patient seen got documented and billed correctly, surfacing only the exceptions. ~70 hours/week of manual cross checking by a three person insurance team down to an exception review; 200+ tests on synthetic fixtures, no PHI in git; live on Cloud Run.
- Centralized Diagnostic Filter. Turning the founder's diagnostic model into a standardized, HYBRIDGE-owned system where AI assists and doctors validate. The owned intake is live on Google Cloud for all three markets, and a Gemini engine now reads the panorex under the founder's own written ruleset (v2 after his review of the first outputs), generating doctor and consultation reports from an internal dashboard.
- Cross-cutting MLOps: one shared ETL module so train/serve skew is impossible. Hot-swap checkpoints via GCS, tag-based Cloud Build CI/CD, OpenTelemetry with strict no-PHI logs.
