I'm
About
I'm a founding team member at Cookiy AI, a Silicon Valley–based realtime user-research platform (funded at eight figures USD, 50k+ respondents and 700+ studies in production).
On the research side, I co-authored AOI (under review at SIGKDD 2026), a multi-agent framework for autonomous cloud diagnosis.
Work Experience
Cookiy is a realtime user-research platform covering the full loop from research plan → recruitment → interview → report, built to let AI agents drive an entire study end to end. I built the realtime multimodal voice interview agent: a cascaded STT→LLM→TTS streaming pipeline with VAD, multilingual turn detection, barge-in and sub-second first response (~300ms LLM / 150–250ms TTFT), plus screen-share vision so the agent can watch and probe real user behavior. I designed the team's agent development harness (contracts frozen into a single CONTRACT.md with CI checks and drift audits — now the team standard) and a pluggable agent runtime built on one event loop and one unified event stream. I exposed the whole platform as 30 MCP tools that passed the official OpenAI Apps review, unified agent/CLI auth via RFC 8414/9728 OAuth discovery with compact tokens ~75% smaller than JWT, and routed 759K+ LLM calls through Cloudflare AI Gateway with ~86% prompt-cache hit rate. I also built the recruitment module from zero: supplier adapters (Prolific / CloudResearch / CINT), a composable screener engine, a concurrency-safe matching & billing engine (wallet dual-write, drift=0), and a guardrailed supplier-selection agent with deterministic fallback.
Co-founded a New Zealand AI startup as one of its part-time tech leads, alongside my Thoughtworks role. Spearheaded a RAG-powered AI bot product aimed at improving enterprise–client communication in traditional industries: designed, built and deployed the retrieval and generation stack in Python with LlamaIndex, set up the CI/CD pipelines, and built a prompt-evaluation pipeline to quantify how prompt variants affect product effectiveness. Departed when the company pivoted to a new business model, having laid the technical groundwork for its next phase.
Enterprise consulting across backend architecture and AI applications, owning features end to end from requirement clarification to production. Built a small-sample RAG customer-support agent PoC on ~10k words of Chinese FAQ (no room for fine-tuning): systematically evaluated Chinese embedding models (m3e / text2vec / E5) for a ~20% precision gain, added BM25 + cross-encoder reranking, built a HyDE-style query store for dual-path retrieval, and set up a reusable RAGAS evaluation pipeline — 90%+ answer accuracy in open testing. Re-architected an Australian streaming platform's metadata backend: Java 21 + Spring Boot + Orkes Conductor video-processing workflows, GitOps delivery with Kubernetes / ArgoCD / Terraform, Kafka + Elasticsearch event-driven monitoring on AWS, and a unified multi-tenant metadata layer for VOD + Live — new business lines onboard within a week, processing efficiency up ~40%, launch cycles cut from hours to tens of minutes.
Skills
Building, every single day
A year of commits across my two GitHub accounts — and how hard I push AI coding tools.
Claude Code Usage Dashboard
I'm a heavy user of AI coding tools — roughly 2B tokens and $1.6k+ in API-equivalent cost burned in the last month alone.
This dashboard is a small GitHub Actions project: a pipeline auto-syncs my daily Claude Code token usage to a public page, split by input / output / cache with per-day cost.
Check out my latest work
I've worked on a variety of projects, from simple websites to complex web applications. Here are a few of my favorites.
Cookiy's open-source skill that turns any coding agent — Claude Code, Codex, Cursor, OpenClaw — into an end-to-end user researcher: AI-moderated interviews, synthetic users, quant surveys, participant recruitment. 900+ stars on GitHub, driving the platform through the MCP / CLI surface I own.
Production voice agent that runs full user-research interviews. Cascaded STT→LLM→TTS streaming with dynamic TTS chunking, VAD + multilingual turn detection, barge-in and silence recovery — end-to-end first response under one second. Screen-share vision lets the agent watch, pause capture, and probe on-screen behavior in realtime.
A pluggable agent runtime built from scratch — single event loop (AgentLoop), unified event stream (AgentEvent), and three swappable contracts (Tool / LlmProvider / SessionStore); no LangChain. On top of it: a Slack-native research agent that parses natural language into tool calls via Gemini function-calling and executes them inside ephemeral E2B sandboxes, with webhook-driven session resume for long tasks.
Wrapped the entire SaaS as an agent-callable MCP server: 30 decorator-registered tools over SSE / StreamableHTTP / stdio, approved through the official OpenAI Apps review. Unified agent & CLI auth with RFC 8414/9728 OAuth discovery and stateless compact tokens (~75% smaller than JWT). All LLM traffic governed through Cloudflare AI Gateway: multi-model routing, rate limiting, cost observability, ~86% prompt-cache hit across 759K+ calls.
Platform-neutral browser-automation layer for AI agents — real Chrome via a Go native-messaging host, driven over JSON-RPC by JS / Python / Go SDKs and a CLI. I built multi-Chrome-profile routing & selection and tab-group management, and reverse-engineered the full native-messaging + CDP stack (benchmarking it against Codex's and Claude's browser layers).
Research & Publications
Alongside product engineering, I work on research about agentic systems — how agents fail, and how those failures become training signals.
AOI: Turning Failed Trajectories into Training Signals for Autonomous Cloud Diagnosis
Under review at SIGKDD 2026 · arXiv preprint · cs.LG / cs.AI
Co-authored a multi-agent framework for automating SRE diagnosis with LLMs: distills expert knowledge into open-source models via GRPO, isolates execution behind a read-write separated architecture, and converts failed operation attempts into training signals — 66.3% success on benchmark tasks, with failure analysis alone contributing +4.8pp.
Get in Touch
Want to chat? Just send me an email and I'll respond whenever I can. I will ignore all soliciting.