Jay Ozer · AI Engineering · San Francisco
I lead teams building production AI for document intelligence and regulated workflows. I like turning messy, manual work into something dependable.
View my resume (PDF)
Selected work
Deed analysis built for the details that matter. Independent extraction teams, neural search, and human review turn a manual title-insurance workflow into a system that learns from its corrections. Live in production with 15,000+ deeds processed; the work now is keeping it reliable as volume grows.
An order surge became a routing problem. I built a two-stage agent pipeline to untangle email threads, identify actionable requests, and send them to the right title-insurance workflow at 95%+ routing accuracy. Then I rebuilt its triage stage around confidence: a small decision model answers first, and a reasoning LLM sees only the cases it is unsure about. Fewer errors, at about a thirteenth of the cost.
Clinical notes that follow the practice's own charting style. PoppyNote listens to the visit, drafts the note, and flags anything unstated for the doctor to confirm. Built so my wife stops staying late.
A new project: an updated version of the DeedIQ running in production, built on Azure Content Understanding with GPT-5.5 to be faster, cheaper to run, and more accurate. Just deployed to production after a full parity test, it ships as a plugin with its own repo and contract, and adds registered-land (Torrens) detection and a Decisions API filter that rejects non-deeds before extraction runs.
One parsing service for every title document. PUDIQ, DeedIQ and 28 other document types sit behind a single endpoint, and every extracted field comes back with a confidence score, so downstream automation makes one call and knows when a person needs to review.
Give it one property address and a swarm of AI agents finds the deed behind it, returning the grantor, grantee and recording date. Benchmarked against known-good deed records and 100 addresses matched with search info, it replaces manual deed searching with a lookup that starts from a single line.
AI agents built to work like title officers and curative specialists: answering emails, checking names, finding additional documents, and powering an auto-typed title product. DeedIQ and DiscoverIQ run as MCP servers and CLI tools the agents call directly. Each bot is built; the work now is the hand-offs between them.
ForageIQ watches screen recordings of people doing a workflow and turns them into working skills AI agents can run. The pipeline is live end to end, recordings in and skills out, and it is rolling out on Resware so recorded title workflows become reusable automations, starting with an order flip.
Tracing and monitoring for Doma's AI agents, built on Claude OpenTelemetry, wired in while it was still in private preview and feeding Datadog and Grafana. A DiscoverIQ dashboard is live, so agent behavior, cost and failures are visible in production and get fixed faster.
I reran Doma's Data Science Challenge with decision models: Jev with GEPA optimization, Decisions API, and tev1, a Qwen 3.5 model fine-tuned on Together AI. It posted the highest score.
A team of specialist agents answers parents' questions for a pediatric dental practice: appointments, insurance, procedures, emergencies. Every answer is grounded in the practice's own pages before it sends, and emergencies skip straight to urgent guidance. The first line of defense on on-call weekends.
A supervised team of Claude Managed Agents writes the Poppy Kids blog. A deep-research agent sources every claim from AAPD, ADA and peer-reviewed work, a creative director makes an on-brand hero image with OpenAI's gpt-image-2, and Dr. Andrea approves topics and drafts over Slack. Nothing publishes without her yes.
Built out of frustration with uploading private files to random compressor sites. FFmpeg.wasm runs in the browser, so video compression happens entirely in the tab and your files never leave it; images get fast WebP conversion. 50 to 90% smaller, with batch support.
A calm iOS kick counter for the third trimester: one tap per movement, a timed session with pause and undo, calendar history, and CSV export for the next appointment. Everything stays on the phone. No account, no analytics; the App Store privacy label simply reads Data Not Collected.
Tap the orb and say it: two ounces of formula, wet diaper, left side twelve minutes. Chloe logs feeds, diapers, nursing and pumping hands-free with on-device Apple Intelligence, talks back to confirm, and turns the day into caregiver handoffs and pediatrician summaries. An iPhone beta ahead of a privacy-only App Store release, named for the person it was built around.
Dictation that never phones home. Hold Fn and speak: local Whisper transcribes on-device, deterministic rules and an optional local LM Studio model polish the text, and one clean copy lands in the focused app. Cloud AI endpoints are refused in code, so audio never leaves the machine: privacy as an architectural fact, not a policy. A subscription is a tool you rent; this one lives with me.
Experience
Led the applied-AI pod delivering production agentic systems for title-insurance automation, partnering with Fannie Mae and Blend.
Built and led a team of six, establishing engineering rituals and modern data practices that scaled company-wide.
Established a unified source of truth and drove company-wide adoption of modern analytics ahead of IPO.
Led data engineering and BI infrastructure processing terabytes of economic data for monetary-policy decisions in a highly regulated central-banking environment.
Patents
Each one started as a stubborn, manual workflow and turned into something that mostly runs itself. The filings are nice; the part I like is that they quietly save people a lot of time.
LLM agents plus neural search with dual-team extraction, cross-validation, and human-in-the-loop review, replacing a brittle, manual deed workflow with a system that learns continuously.
Hybrid automation across fragmented county portals, with domain-specific parsing, template-driven scripted browsing, AI fallback, and vision verification that every document matches before delivery.
Fully autonomous retrieval for county clerk portals: it discovers the path and turns successful runs into reusable skills, while agentic reasoning corrects bad input data and adapts to portal changes with no human in the loop. Now in production, shipped with the Expansion plugin.
Working with people
A decade in, the part I care about most isn't the systems. It's the people I build them with. I've hired engineers from zero, sat through the boring reviews, and learned that patience ships more than cleverness. No heroics, just steady, honest work.
Open source & local tools
Python-first coding-agent Kanban that turns a large goal into scoped stories, runs one worker at a time, and keeps an auditable review trail, with gates that stop a worker from editing the test it is graded by.
View repoA calmer Gmail inbox powered by Jev that runs entirely on your computer. Evaluates priority, categories, and reply signals with human-in-the-loop label previews, verification receipts, and undo.
View repoHelp me be understood. Turns notes, docs, repos, demos and recordings into source-grounded product presentations delivered in your own voice, which the audience can interrupt with questions.
View repoA marketplace of installable AI-agent skill kits, including workflow orchestration, YouTube automation and video understanding tools.
View repoBedtime stories told in a parent's voice. The live site is a beta front door, while the iOS app moves toward a privacy-only local version for saved voices and stories.
Visit site14 standardized tools exposing my own bio, experience and CV over the Model Context Protocol. Built in under an hour.
View repoFixed a critical duplicate-download bug, hardening download handling across CDP, JS-fetch and polling paths.
Production read and write MCP server for OpenAI's AgentBuilder, with dual Graph plus org-account auth. Powers IntentIQ.
Edit-before-upload document prep for Voiceflow RAG. The community adopted it; Voiceflow shipped the idea natively.
Ultra-fast LLaMA Guard guardrails for Voiceflow via Groq, real-time content safety with minimal latency.
A transparent, PRD-driven Bash loop for fresh-context coding iterations across Claude Code, OpenAI Codex and Kimi CLI.
View repoSnap a snack photo and Dr. Hawley, a wisecracking molar, roasts it in a kid-safe comic strip. Deterministic dental-risk scoring stays in charge; the AI only does the jokes and the art.
Visit siteAudits any public website for agent-readiness and speed using Is Agentic, Ora, and Google PageSpeed Insights (mobile and desktop) without API keys, then gives prioritized fixes and before/after verification.
Try AgentReadyWriting
Two local inference engines, one Mac, the same private clinic-data loop. Splash generates up to 2.4x faster and answers repeat prompts about 4x sooner, but it serves one request at a time.
Generation got cheap. Verification didn't. Four months on a project where every individual sped up and the whole thing slowed down.
The one step I kept for people, because AI couldn't do it, came back generated by a chatbot. On automating versus offloading, and who still checks the ground truth.
Ask Jay
A retrieval-augmented agent grounded in my notes, projects and patents. Ask how a system is put together, how I think about a hard problem, or just say hi. It answers with citations.
About
I grew up in Istanbul, played competitive sports, and came up through engineering before finding my way into data and AI. Sport taught me what I still lead by: you rarely win by being exceptional at one thing. You win by being above average at many, and combining them well.
Music makes the same point. There was a time when a musician could write one great song and tour the world on it. That era is gone. The artists creating moments now are DJs: they collect the best tracks and mix them into cohesive sets, each one built for the room in front of them. Building with AI is exactly that. It is not music production, chasing one perfect track. It is DJing: curating the best pieces and blending them into something that moves people.
Home is San Francisco, with my wife Andrea, our daughter Chloe, and our golden retriever Hawley. When I'm not shipping, you'll usually find us out on a long walk, or at home playing records for the family.
Say hello
I love comparing notes with people building thoughtful AI, a hard problem, a half-formed idea, or just a good conversation. If you're making something interesting, I'd like to hear about it.
