I'm a Senior AI Engineer at MoMo, Vietnam's largest digital wallet, on the Business AI / Credit Tech team. I design and ship the AI systems behind consumer lending — from data pipelines and decision models to LLM agents running in production under real latency, cost, and compliance constraints.
My work sits where agent architecture, ML engineering, and distributed systems meet. I care about the unglamorous part of AI: making systems reliable, measurable, and cheap enough to run every day — not just impressive in a demo.
- 🏦 Building AI for fintech & consumer lending at scale
- 🤖 Designing LLM agents with memory, evaluation, and self-improvement loops
- 🎙️ Shipping real-time voice agents with low-latency speech pipelines
- 📍 Based in Ho Chi Minh City, Vietnam 🇻🇳
Designed a meta-agent that analyzes graph-based LLM agents, proposes changes, and evaluates them in a closed loop (setup → optimization → human-approved exit). Turns agent tuning from manual prompt-tweaking into a repeatable, auditable engineering process.
Architected a tree-structured memory layer with typed nodes (episodic, semantic, goal, topic, file) and LLM-based retrieval routing, so agents recall the right context instead of the most similar context. Includes an explorer UI for inspecting and debugging agent memory.
Building speech-to-speech agents across cascaded (ASR → LLM → TTS) and end-to-end pipelines, focused on conversational latency, voice-activity detection, and natural barge-in handling.
Engineered core components of MoMo's lending stack, including a data abstraction layer that unifies feature access across credit products, feeding underwriting and decisioning models serving [X]M+ users.
| Area | What I'm working on | |
|---|---|---|
| 🎯 | Agent Reliability | Evaluation harnesses, trajectory analysis, regression testing for agent behavior |
| 🧠 | Agent Memory | Structured long-term memory, retrieval routing, memory consolidation |
| 🧩 | ML Systems | Training and inference pipelines, recommendation and decision systems |
| ⚡ | LLM Infrastructure | Scalable serving, GPU utilization, observability, cost optimization |
| 🌐 | Distributed Systems | QUIC, P2P networking, low-latency architectures |
📏 Measure before you optimize. An agent without an eval suite is a demo.
🛡️ Design for failure. Production AI fails in ways benchmarks don't show — build the guardrails first.
⏱️ Latency and cost are features. If it's too slow or too expensive to run daily, it doesn't ship.
🤝 Open to conversations about agent reliability, ML infrastructure, and AI in fintech.