I'm a Full-Stack AI Engineer who ships products to real users, not demos. Five-plus years across AI, data, and software engineering — from NLP research labs in Bangkok and Paris to zero-to-one startups at Station F, to founding my own studio.

Today I'm building Ekkhara, a self-funded AI ventures studio, while engineering production systems at the intersection of health tech, regulatory compliance, real-world evidence, and telecom data infrastructure.

• 🏗️ I architect first, then build. Clean / Hexagonal architecture, API-first design, real tests, CI that stays green.
• 🤖 My specialty: RAG systems, production AI agents with observability & failure detection baked in, cloud data pipelines, 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.

Real products, real users, real moats. Each one shipped end-to-end.

The one that's genuinely shipping to real Myanmar learners. A TOEFL speaking & English-practice bot that runs entirely inside Telegram — so learners can train for the computer-based TOEFL despite the country's internet restrictions, no VPN needed. I built the full stack: Python/FastAPI services, LLM-driven speaking feedback & calibrated scoring, and the conversational UX.

Python · FastAPI · LLM · Telegram Bot API — ▶ Live → @SpeakProofTOEFLBot

Built & live on GCP Cloud Run. Mistral OCR over French lab reports + deterministic LOINC biomarker classification + FHIR R5 audit trail for personalized supplement guidance. The moat is the data + validation pipeline, not the LLM.

Python · Next.js · FastAPI · PostgreSQL · Mistral OCR · FHIR R5 — 73 tests · 86% coverage · Haleon @ VivaTech 2026

Built & live on GCP Cloud Run. A 60-second daily check-in habit loop with Mistral-Vision meal analysis and a longitudinal Vitality Score that compounds over 30 days. Retention is the moat.

Python · Mistral Vision · GCP Cloud Run — 64 tests · Nestlé @ VivaTech 2026

Production-grade platform for vaccine researchers: PICO protocol builder, PRISMA screening pipeline, RoB 2 / ROBINS-I assessment, meta-analysis forest plots, real-time CRDT collaboration, and FDA / EMA / CDISC regulatory exports.

Next.js 16 · React 19 · TypeScript · Supabase — 76 API routes · 27 DB tables · 1,400+ tests · ▶ Live

Probabilistic forecasting and stochastic optimisation for battery-flexibility decisions on a real-time AWS lakehouse. Streaming ingestion → feature store → ML serving, fully orchestrated.

AWS · Apache Iceberg · Kafka · dbt · Airflow · MLflow

Event-driven Call Detail Record ingestion, rating, and reconciliation pipeline — the kind of system that bills real mobile traffic. Idempotent, replayable, observable.

Java 21 · Spring Boot 3.5 · Kafka · MySQL · MongoDB · Docker

A ReAct agent built observability-first — a failure taxonomy and evaluation harness that catches where agents break, with live SSE streaming of reasoning traces. This is how I think production agents should be built.

Python · FastAPI · Next.js 16 · PostgreSQL · ReAct · Groq · SSE

End-to-end CSRD / ESRS sustainability-reporting reference implementation — Snowflake in the cloud, DuckDB locally, dbt transformations, Airflow orchestration, LLM-assisted disclosure mapping.

Snowflake · DuckDB · dbt · Airflow · Claude · Mistral

Human post-edited English→Myanmar instructional MT corpus, an NLLB-200 fine-tune benchmark, and a novel Instruction Faithfulness Score for evaluating low-resource translation.

Python · NLLB-200 · HuggingFace · PyTorch

More on pseonkyaw.dev — Diameter Credit-Control (Gy/RFC 4006), SMPP Gateway, Mobility Pulse (TimescaleDB + PostGIS + H3), BCBS 239 Lakehouse, and more.