Hello, I'm Long.
AI-Native Product Engineer
LONG TRỊNH
about
01I'm Long Trịnh, an AI-native product engineer based in Nha Trang, Vietnam. I build agent-first products and operate them in production: an MCP flight tool whose watchlist found a real €1,707 booking, a hotel PMS that runs my own front desk every day, and a fine-tuned local model published with its failing exam score.
I operate what I ship — including the 10-room hotel my software runs, where guest ID registration is a legal requirement handled by my own pipeline every day. If a flow is confusing, I hear about it the same day.
Agent tooling: Python, Pydantic, MCP. Products: Next.js, React, TypeScript, Tauri, Rust, Flutter, SQLite, Postgres. Operations: n8n, on-device OCR, fine-tuned local models.
services
02Tools AI agents can actually trust: strict typed contracts, protocol hygiene, and honest failure modes.
I design the tool surface, lock the schema, and keep the protocol clean — so an agent can act on results instead of scraping prose.
From spec to production across web, desktop, and mobile — including knowing where the AI shouldn't be.
I scope the product, write the specs, direct AI coding agents under executable guardrails, and gate every release on verification.
Automation that handles sensitive data responsibly — on-device where possible, with one layer owning every write.
I wire orchestration, data, and monitoring so mandatory, repetitive work runs itself without sending private data anywhere it doesn't need to go.
Every project here answers the same question: which kind of intelligence does this problem need — a frontier LLM, a fine-tuned local model, deterministic rules, or none at all — and can it survive production?
An agent-first MCP server: one tool, a strict Pydantic contract, and a watchlist that found a €1,707 booking.
Built when war pushed Europe-Vietnam fares to record highs. The agent holds the fuzzy intent; the tool returns typed offers with honest caveats about what it didn't search — and the watchlist caught the route that saved about 43% on real tickets home.
A production hotel PMS written mostly by AI coding agents under executable guardrails — running my own front desk every day.
No-float-money gates, append-only ledgers, and a command boundary keep stochastic programmers safe. The app ships its own on-device fine-tuned agent with zero write access.
A TikTok channel produced by an agent pipeline.
A faceless Vietnamese finance-explainer channel: 3.9K followers and 622.7K views in its first weeks, 80.2% from the For You feed, at ~$1 of API cost per video. Built on the open-source OpenMontage platform; my layer is the editorial DNA — a custom Remotion doodle engine, Vietnamese voice skills, and QA gates and human review. One topic goes in, one narrated video comes out: nine stages run on their own and stop for a human exactly once.
A fine-tuned local model that behaves like a bilingual hotel receptionist — published with the exam it failed.
A synthetic-data factory of 9,564 curated Vietnamese/English conversations across five versions, Gemma fine-tuned with LoRA, and an honest Fixed30 eval shipped alongside the quantized GGUF — because at a front desk, knowing when not to answer is the job.
An LLM ops layer where the LLM is the fallback, not the brain — n8n orchestration with Airtable as the single source of truth.
400+ guest IDs read into a CRM from one double-click: 68% fully on-device, 84% with no human touch, and a Python layer owning every write. Privacy-first by architecture.
A coaching engine that is deliberately not an LLM.
A weight-loss running coach born from my own 98→75 kg run. Coaching advice touches health and trust, so the plan engine I'm building is specced as deterministic rules — same inputs, same advice, no hallucinated training plans. AI stays where variance is acceptable: reading run screenshots, the part that ships today. Live prototype, single user: me.
The unglamorous 80% of a generative AI product: queues, keys, storage, and accounts.
Background job queues keep the UI responsive through slow generation, bring-your-own API keys are encrypted at rest, and panels land in per-user asset storage — shipped as both a local app and a hosted SaaS on Vercel.
Zero AI, on purpose — a privacy-first family tree app for Android that works fully offline.
A custom tree canvas, Vietnamese kinship modeling, local SQLite storage, and file-based backup and restore. No accounts, no server, no analytics. Live on Google Play.
contact
04Building agent tooling or an AI-first product? I work with EU teams — afternoons in Vietnam are mornings in Europe, so we share half a working day live. Email me; I reply within one business day.
closing
Everything above is public, shipped, or running a real business today.
© 2026 Long Trịnh