English · Tiếng Việt
data pipeline
lakehouse
streaming
RAG
text-to-SQL agent
business chatbot
Data Engineer with 2 years of experience, based in Hà Nội. I build data pipelines (Airflow, Kafka, Spark, dbt) and AI agents that answer on top of that data (RAG, text-to-SQL, tool calling).
taxi-lakehouse-ai-agent — ask NYC taxi data in plain Vietnamese
- Problem: let people who don't write SQL query the data in natural language, without the agent being able to run a dangerous statement or read outside the allowed data.
- Approach: Airflow → MinIO (Bronze) → dbt (Silver → Gold star schema) → DuckDB. FastAPI agent: intent analysis → rule-based planner or LLM-generated SQL →
sqlglotguardrail (SELECTonly, cataloged Gold tables only, valid columns and joins) → read-only execution → self-checks → answer with SQL and trace. - Results (
gpt-4.1-mini, measured 2026-06, reproduction commands inbenchmarks/):
| Axis | Result |
|---|---|
| Guardrail — 24 attack queries (DML/DDL, injection, non-Gold tables, file reads…) | 24/24 blocked, 0 false positives |
| Spider dev — 1,034 questions, official evaluator | EX 81.1% (GPT-4 zero-shot: 72.3%) |
| 96-question custom set on the taxi domain | 78.1%, up from 40% through agent improvements |
tradewatch — streaming that stays correct when things break
- Live trades from the Binance WebSocket → Kafka (KRaft) → Spark Structured Streaming → Postgres (idempotent sink) + Parquet; public dashboard on Cloudflare Workers + D1.
- The repo is organized around reproducible experiments (
make exp-NN), each answering one question: does a job dying mid-micro-batch lose or duplicate records? Is a record arriving 3 hours late dropped, or does it correct the old result? Does exactly-once come from Spark or from the sink? - Dirty data is generated by a fault injector (late, duplicate, null, schema drift); design decisions are recorded as ADRs.
- realtime_fraud_detection — real-time transaction fraud detection: Kafka, Spark Streaming, PostgreSQL, Elasticsearch, Grafana.
- youtube_analytics — YouTube API → Snowflake, dbt Bronze/Silver/Gold models, orchestrated with Airflow (Astro).
- Prestige DMC — production travel platform: Express + PostgreSQL backend (82 endpoints, 19 tables), deployed with Docker Compose, GitHub Actions, GHCR, Traefik and Cloudflare Tunnel.
🌐 Open-source — Tencent/WeKnora · ⭐ 30.7k
Tencent's open-source RAG / agent platform. I fix bugs and propose features upstream. As of 2026-09-28: 2 PRs merged, 2 in review, 10 issues.
| PR | Status | What |
|---|---|---|
| #3520 | ✅ merged | docreader: a base64 payload containing non-ASCII characters failed the whole document parse |
| #3673 | ✅ merged | Restored inner padding on form popups after a style consistency pass (#3672, reported by me) |
| #3522 | 🔍 review | The agent emits the knowledge references it retrieved during a turn, so the UI can show sources |
| #3187 | 🔍 review | Outline datasource connector (collections and nested documents), +2,278 lines |
Notable issues: carry the user's own language into the prompt on IM channels (#3315), distil recurring questions into reviewed FAQ entries (#3316).
- Data: Python, SQL, Airflow, dbt, Kafka, Spark, PostgreSQL, Snowflake, DuckDB, AWS S3, Playwright, Scrapy
- AI / agents: RAG, text-to-SQL, tool calling, MCP, OpenAI API, n8n
- Infra: FastAPI, Docker, Jenkins, GitHub Actions, Traefik, Grafana
HUST — Mathematics & Informatics, 2022 – present
- Email: chuquangvinh.work@gmail.com
- LinkedIn: linkedin.com/in/chu-vinh