Matt Kerkstra

Operating at Staff scope across a 45-person engineering org: own production LLM products, the Kubernetes ML platform, and source-system integrations.

• ›Co-led AI product delivery for Canon, replacing fixed clinical-note templates with a block-based system that learns each customer's structure from examples. Built the evaluation and evidence-gated rollout around deterministic diffs, LLM-as-judge over semantic differences, and clinician calibration. Cutover proceeded customer by customer, and clinician edit rate per section fell from roughly 60% to 40%.
• ›Shipped the first production clinical-note LLM system with two teammates in 16 days across 3 services: graph-based orchestration with Postgres checkpointing, BAAI/bge-m3 embeddings + Milvus retrieval, dual-detector PHI anonymization (Presidio + LLM), SSE streaming, and an admin review UI. Onboarding shipped as an agent skill for iteration speed, with the hardening plan written down up front.
• ›Built the AI evaluation and feedback loop with the voice and product teams: clinician edits attributed to the block that produced the text, per-block performance statistics, and edit-pattern dashboards. Turned prioritization debates into measured user behavior and gave clinicians the deciding vote on what ships next.
• ›Own the Kubernetes ML platform the org deploys on: Istio ambient mesh, ArgoCD app-of-apps with Kustomize overlays, KServe + vLLM self-hosted model serving (embeddings, ASR), Langfuse for LLM observability, Milvus for vector search - 13 namespaces / 9 services. Built with the platform team; new services ship with a Helm chart and an ArgoCD app, cost stayed flat, and six engineers contribute regularly.
• ›Designed a production shadow evaluation with the voice team: both diarization variants generated on sampled real sessions, the delivered note never waits, and go-live is a reviewer-gated config change. The design answers "does speaker attribution improve the output?" with production traffic and zero patient-facing exposure.
• ›Led the MongoDB → PostgreSQL/PostGIS migration of 100M+ clinical analysis records with the data and CV teams; heavy queries dropped from >5s to <500ms. Own the chain from schemas through Python treatment-recommendation algorithms to TypeScript integrations across 10+ versions.

  compounding arc

  postgis seed · 3yr cascade

• ›Built the end-to-end voice-to-clinical-note product: offline-resilient React/browser capture, BullMQ async processing, transcription, structured LLM summarization, and practice-management-system context. More than 60K TypeScript LOC across client and API.
• ›Delivered the unified-appointments backend behind the largest dental services organization (DSO) contract in the United States - vault practice search, dual-mode practice support, legacy PMS integration optimizations.

  huddle view

  one query · imaging × pms

• ›Run the bi-weekly Backend Guild (18+ months): typed query patterns (Kysely + footgun-prevention bots) adopted org-wide, AI-assisted developer tooling, architecture decisions written down as ADRs.

• ›Rebuilt Node.js + Stripe billing to support 2.5× YoY revenue growth while keeping payments under 200 ms.
• ›Implemented real-time fraud and sanctions checks that shut down illicit GPU crypto-mining from embargoed regions, reducing chargebacks >50%.

  request filter

  chargebacks −50%

• ›Introduced bundle splitting, CDN routing, and feature flags, shrinking mean deploy time from 15 min to 6 min for 40+ engineers.
• ›Drove Redux → React Query migration, reducing cold-start data fetches 40% and bundle size 20%.

  time to interactive

  redux → react query migration

• ›Converted a 20-year-old version-control system for F&I forms from VB6/SQL to COBOL + Pick BASIC - responsible for distributed delivery, usage tracking & billing, and integration with F&I systems. Assumed lead role two weeks after onboarding and delivered on schedule.

  FORM-VCS v2.3

  vb6 → cobol + pick basic