I build agent workflows as explicit graphs — then define how they're checked, trace every step, and observe the results.
Define → Build → Trace → Observe. Every project pairs what was built with how it's evaluated.
I'm a Full-Stack AI Engineer who ships products to real users, not demos. Engineering since 2021, AI/ML since 2023 — from NLP research labs in Bangkok and Paris, to zero-to-one startups at Station F, to agentic pipelines as a freelance AI engineer.
Most recently at Hoora Games (SAS EASYWIN) I owned the evaluation and quality-gating layer of an agentic game-generation pipeline built on Mastra / TypeScript. Today I focus on agent graphs you can measure: routers, evaluators, tracing and observability.
- 🏗️ I architect first, then build. Clean / Hexagonal architecture, API-first design, real tests, CI that stays green.
- 🤖 My specialty: agent workflows and routers, evaluation harnesses (deterministic detectors, mutant testing, LLM judges), RAG / GraphRAG, observability, and LLM fine-tuning.
- 🎓 Dual Master's in Data Science — Télécom SudParis (Institut Polytechnique de Paris) 🇫🇷 & Asian Institute of Technology 🇹🇭.
- 🌏 Yangon → Bangkok → Paris. Social scientist turned engineer — communication and cross-cultural instincts are part of the toolkit.
Each one pairs what was built with how it's evaluated — and says where the evaluation stops.
🛰️ Faultline NOC — Agent Router + Evaluation Harness
One router in front of three specialist agents, scored before it can touch a network. A typed RoutePlan contract (ordered steps, per-step context, which step may write and whether it needs confirmation), a keyword baseline and two Claude routers answering via a forced tool call — on a seeded, simulated 5G SA core. The harness is itself tested: 4 mutant routers each carry one planted defect and must trip their own detector. CI replays every recorded model call with no API key.
Python · pydantic · mypy strict · pytest · Anthropic tool use — Claude Haiku 4.5 & Sonnet 5: 45/52 vs keyword baseline 40/52 · ▶ Design note · Project page
🎮 Agentic Game Generator — Mastra Workflow Graph + Two-Tier Evaluator
Generates single-file HTML5 mini-games through an explicit workflow graph: plan → generate ⇄ verify (repair loop capped at 3) → finalize | reject, with typed Zod I/O and every failure attributed to its step. Checked by 24 deterministic contract rules, a Playwright runtime probe, a cited LLM judge, and a language check. Key finding: 9 of 20 games that pass every static rule still fail at runtime.
TypeScript · Mastra · Zod · AI SDK · Playwright · Vitest — 20/20 prompts produced a game · replayable from committed cassettes · ▶ Live demo (replayed model calls)
📡 AgentPulse — Real-Time Observability for Agent Runs
Traces, tool calls, tokens, cost, latency and errors — streamed live from a swappable event bus (memory / Pub/Sub / Kafka drivers) into a span waterfall and a 60-second rolling p50/p95 aggregator. One Zod event contract, checked at ingest and in the browser. Demo traffic is simulated.
TypeScript · Fastify · Vue 3 · Kafka · GCP Pub/Sub · SSE · GraphQL — 76 tests · ▶ Live demo
📚 WikiHow-MY — English→Myanmar MT, Fine-Tune & Human Study
~10K human post-edited pairs with article-disjoint splits, an NLLB-200 600M fine-tune (weights on HF), a 4-system benchmark (chrF++, spBLEU, COMET, MetricX-24), and a 9-rater human study — which showed my own proposed metric fails (r = 0.084). The negative result is reported.
Python · Transformers · NLLB-200 · COMET · MetricX-24 — +5.63 chrF++ in-domain · +4.33 on FLORES+
🔍 AgentProbe — AI Agent Failure Taxonomy & Eval Harness
A ReAct agent written from scratch, with a deterministic failure taxonomy, a 52-case evaluation harness and live SSE streaming of reasoning traces.
Python · FastAPI · TypeScript · SSE — Recorded benchmark: Claude Haiku 4.5, 42/52
GraphRAG stages on an interactive canvas with a retrieval eval (vector-only vs graph-expanded), and a Vue Flow studio for composing agent workflows with step-level run visibility.
Vue 3 · VueFlow · Node/TS · Mastra · BAML — ▶ KG-RAG demo (mock mode) · ▶ AgentCanvas demo (mock mode)
💉 VaxEvidence — Real-World Evidence Platform
PICO protocol builder, PRISMA screening, RoB 2 / ROBINS-I assessment, meta-analysis forest plots, real-time CRDT collaboration, and FDA / EMA / CDISC exports.
Next.js 16 · React 19 · TypeScript · Supabase — 76 API routes · 27 DB tables · 1,400+ tests · ▶ Live demo
⚡ EV Charging Copilot · 🪪 FaceProof · 🗣️ SpeakProof
A tool-grounded RAG assistant where every station, price and route comes from a tool call · face verification + liveness detection measured on the LFW protocol · a TOEFL speaking-practice bot inside Telegram used by learners in Myanmar.
Data & back-end range: GridFlex (AWS grid lakehouse) · CSRD Lake (Snowflake + dbt) · CDR Pipeline & Diameter Credit-Control (Java 21 · Spring Boot 3.5 · Kafka) — more on pseonkyaw.dev.
Building at the frontier of AI, data, and product — from Station F to the rest of the world.
Open to mid-to-senior roles & collaboration — AI Engineer · ML Engineer · Applied AI / Evals.
📫 Always happy to talk AI, data, or building something ambitious.