Analytics · Data Platform · Machine Learning · GenAI

Trần Quốc Việt (Charlie) — Data Scientist (Credit Scoring & ML) @ Cake by VPBank (digital bank, HCMC).

I ship credit scoring end to end — from curated marts and feature design to AutoML-style training (CakeAutoML), validation (OOT / stability), monitoring (Gini, KS, PSI), and production scoring APIs. I also build GenAI systems — agent harnesses, memory & context engineering, hybrid RAG, agent skills / tool contracts, eval & tracing — and the data platform underneath (Airflow, dbt, BigQuery, streaming/batch). I care about observability, least-privilege data access, and systems risk teams can operate in regulated environments.

Credit scoring & risk (primary)

• End-to-end: feature design, CakeAutoML training, train–val–OOT, production scoring APIs.
• Models & methods: scorecards, WOE / IV, XGBoost, calibration, decision thresholds, reject inference, champion–challenger.
• Operations: monitoring (Gini, KS, PSI), drift / stability, Feast-style feature views.

GenAI & agents (production-oriented, SOTA patterns)

• Agent harness — orchestration, routing, tool loops, human-in-the-loop; multi-agent workflows (planner / executor / critic).
• Memory — short- vs long-term stores, cross-session recall, summarization & compaction; patterns like Mem0 and durable user/project state.
• Context engineering — token budgets, prioritization, structured prompts, retrieval grounding, history compression without losing task state.
• Skills & tools — agent skills (SKILL.md playbooks), MCP / function-calling contracts, idempotent tools, discoverability for LLM callers.
• RAG — hybrid retrieval, reranking, eval harnesses, guardrails; NL→SQL / analytics assistants.
• Ops — Vertex AI / gateway routing, Langfuse-style tracing; sandboxed execution where agents touch code or data.
• BeGuru AI — FastAPI / AgentOS runtime, memory planes, artifacts on disk, design system for agent-built UI.

Data platform

• Airflow DAG quality (idempotency, SLAs), dbt tests & docs as contracts, BigQuery, streaming + batch ELT.
• Credit feature EDA: strength, stability, cohort documentation for backtests and handoffs.

More on charlieviettq.github.io when posts are ready.

_{Generated by GitHub Actions on branch output; rerun workflow if the image is missing.}