AI agent engineer — production LLM agents that drive real browsers.
Montréal, QC · building agentic systems at Deck.co since April 2025
Deck is an AI-native data platform: LLM agents log into third-party portals on a customer's behalf and return structured data and documents through an API. I work across the agent runtime — the loop itself, the browser tooling it calls, what it costs, whether it succeeds, and whether we can prove why. Three teams: reliability, throughput to new sources, and the v2 platform.
This code is proprietary, so none of it is linkable. The open-source work below is the public analogue of the same expertise — agent tracing, failure classification, provider benchmarking — written from scratch, on my own time, against the same problems.
Playwright traces tell you what the browser did. Agent logs tell you what the model said. Neither tells you, across ten thousand runs, what fraction failed because of bot detection versus stale selectors versus bad credentials. That gap is what these repos are about.
Local-first toolkit for building, debugging and observing web-browsing agents. Four composable repos, MIT.
Trace schema and Python tracer for agent browser sessions. Replay a run step-by-step; classify failures as agent decision vs. site change vs. antibot vs. credentials. On PyPI.
Breakpoints for agent runs. Drops into pdb — and pauses the real browser — the moment a semantic event fires: antibot detected, login failed, MFA prompt, or your own predicate.
Zero-dependency single-file HTML viewer. Drag in a trace, get the step timeline, screenshots and an antibot heatmap.
Agent-native browser. Shipping today as an MV3 extension that captures real sessions into the trace schema; an RFC for the engine-level fork it wants to become.
One interface over autonomous agent-browser providers, with routing, fallback, a confidence gate that catches an engine claiming success on a login-failed page, and a graduation log.
The same authenticated task — log in, download the statement — run head-to-head across three AI-browser providers. Success is a real PDF on disk, not a claim.
A browser capability where the method name is the trace event type — bridging compile-time effect declarations and run-time semantic events. Read a signature, know what its trace will contain.
A Montréal café guide assembled entirely out of text messages. You text an agent, it saves the place or searches what others added; photo-first, any language, self-correcting. Drafts stay private and publish an hour after you stop texting.
screenshot · running locally
Montréal's community pothole map — report one, vote the worst up, watch them get filled. Cone markers carry their vote count, filled holes turn green, and the whole app talks to one store interface. Leaflet, no API keys.
screenshot · running locally
Live-streams a coding-agent session to a web page over SSE — prompts and tool activity, redacted by default (keys, tokens and JWTs scrubbed; never file contents). Optionally type back into the session from the browser.
screenshot · synthetic session, real UI
Lead-hunter for people selling agent work. Polls the repos where the buyers congregate, has a model score budget signal, relevance, specificity and archetype, and delivers only what clears the threshold — to a webhook you own, not a bot you have to host.
diagram
Turns "give me every X in area Y with contacts" into a sourced CSV. The value is reconciliation plus live verification, not scraping: aggregators copy each other's stale rows, and the live browser pass corrected paid directories by more than 1.5 stars on real businesses. Every gap is marked unknown rather than filled in.
diagram
Your phone rings a few minutes before a calendar event and a voice tells you what's starting. One cron worker, nothing to keep alive, no inbound webhooks — and a key-value note per event so it never double-calls you.
diagram
Word-synced kinetic typography and deep-glow motion graphics from PIL and ffmpeg — no After Effects — plus three repair recipes: fill a dead gap with narration-synced animation, remove a person from a screen recording, and cut a 16:9 demo into vertical.
diagram
Agent-to-agent dating: two agents negotiate on their humans' behalf and hand back a plan. An experiment in what happens to a conversation when the other side of it is also an agent.
diagram
Available for contract work alongside the day job — remote from Montréal, or on site if you're here. Fixed price and fixed scope, so you know what you're buying before you commit. The scoping call is free and I'll tell you if it's not worth doing.
I attribute your spend per tool, per turn and per model, ship the cuts that survive an eval, and hand you the attribution script so you can keep doing it after I leave. On the production fleet at work this took out a ~$40k/month run-rate with no drop in success rate.
One workflow end to end: log in, navigate, extract or download, with retries, failure classification and traces you can actually read. Playwright or CDP, MCP tools, your model of choice — plus an honest report on what it still can't do.
MCP servers, agent skills, and bots that live where your team already works — Slack, Notion, Linear — deployed with proper secret handling and shared state. I run a fleet of eight of these at work.
Static site, Stripe Checkout, scheduling and transactional email, live on your domain — plus a plain-language manual so you can change anything yourself with an AI assistant instead of paying someone every time a price moves.
Evals, fine-tuning and distillation, scraping that keeps breaking, data enrichment, video and motion graphics, a prototype you need by Friday. If it's AI and it has to survive production, describe the problem and I'll tell you honestly whether I'm the right person for it.
→ sunwoongha1@gmail.com — tell me the problem, not the spec.
Two reasons this is on the page: shipping a product needs the same nerve as filling a room, and a phone that doesn't scare you is a real engineering asset when you have to go ask a stranger why their portal broke.
Agents & LLMs — Claude Agent SDK, Bedrock Converse, MCP servers and tools, sub-agent architectures, prompt-cache economics, structured output, LLM-as-judge, fine-tuning (Qwen on Vertex AI), token-cost attribution.
Browser automation — Playwright, CDP, Chrome extensions/MV3, bot-detection and WAF signals, proxy strategy, agent-browser providers (Browser Use, Browserbase/Stagehand, Anchor, Bright Data).
Languages — Python, TypeScript/JavaScript, C# (production), Java, Swift, C++.
Platform — GCP (Cloud Run, Pub/Sub, Vertex AI, Secret Manager, BigQuery), Azure (AKS, Blob), AWS (Bedrock), Cloudflare (Workers, Pages, KV, cron), Docker, RabbitMQ, Redis, Terraform.
Observability — OpenTelemetry, Grafana, Loki, Prometheus/PromQL, LLM trace tooling, Metabase, BigQuery.
Ways of working — Linear, agent-assisted development end to end, and a habit of building the measurement before the fix.
| Repo | What it is |
|---|---|
| agent-trace | Trace schema + Python tracer + CLI replay for agent browser sessions |
| agent-debugger | pdb breakpoints on semantic agent events |
| agent-viewer | Single-file trace viewer with antibot heatmap |
| agent-browser | MV3 capture extension + RFC for an agent-native browser |
| agent-stack | Umbrella for the four above |
| agentic-engine | Provider-agnostic agent-browser engine, routing + fallback + confidence gate |
| browser-bakeoff | Three AI-browser providers on one authenticated task |
| zero-browser | Effect-declaration ↔ trace-event bridge |
| claude-live | Stream a coding-agent session to the web over SSE |
| problem-bot | LLM-scored lead-hunter over GitHub issues |
| data-enrichment-skill | Sourced, live-verified local-market CSVs |
| mtlpothole | Community pothole map for Montréal |
| sf-trip-skill | Structured city reconnaissance dossiers |
| a2a-date | Agent-to-agent dating experiment |
| matcha-scout | SMS-native city guide (private) |
| gcal-call-reminders | Phone call before your calendar events (private) |
| mograph | Kinetic-typography + video repair recipes (private) |
| better-claude-rc | Drive a desktop coding agent from an iPhone over Tailscale (private) |
| teach-yourself-cs | Self-study notes |
| portionify | Grocery portioning app (Svelte) |
| openplayer | Open-source music player (React) |
| halbert | Node scraper for a copywriting archive |
| covid-tweets-data-crawling | Tweepy + Firestore ingestion pipeline |
| stay-hydrated | Arduino tilt-sensor hydration nag |
| simplecal | Android calendar |