Credit Decisioning · GenAI Agents · Data/ML Platform

Trần Quốc Việt (Charlie) — Data Scientist building Fintech AI/ML Systems @ Cake by VPBank (digital bank, HCMC).

I build AI/ML systems for fintech decisions — from credit decisioning and production scoring to GenAI agents, customer automation, analytics agents, and the data/ML platform underneath. My deepest domain proof is credit scoring: CakeAutoML, label preparation, feature design, train/validation/OOT workflows, calibration, monitoring (Gini, KS, PSI), and handoff into production decision systems. I care about observability, least-privilege data access, and AI systems that risk and operations teams can actually run.

Credit Decision Science (domain depth)

• End-to-end scoring workflow: label preparation, feature design, CakeAutoML training, train–val–OOT, validation, and production scoring handoff.
• Decisioning methods: scorecards, WOE / IV, XGBoost, PD bands, calibration, cutoffs, decision thresholds, reject inference, and champion-challenger.
• Portfolio operations: monitoring (Gini, KS, PSI), drift / stability, model readouts, and Feast-style feature views.

GenAI & Agentic Systems (production-oriented)

• Multi-agent workflows — orchestration, routing, tool loops, supervisor patterns, and human-in-the-loop controls.
• RAG & analytics agents — hybrid retrieval, reranking, NL→SQL, metadata discovery, charting workflows, eval harnesses, and guardrails.
• Agent skills & tools — SKILL.md playbooks, MCP / function-calling contracts, idempotent tools, and discoverability for LLM callers.
• Context & memory — token budgets, prioritization, retrieval grounding, history compression, and durable user/project state.
• Ops — Vertex AI / gateway routing, Langfuse-style tracing, monitoring, and sandboxed execution where agents touch code or data.

Data & ML Platform

• Airflow, dbt, BigQuery, Kafka, Doris, and streaming + batch ELT for analytics, automation, and scoring workloads.
• Feature-store patterns with Feast, production scoring workflows, DAG orchestration, data contracts, and repeatable model pipelines.
• Credit feature EDA: predictive strength, stability, cohort documentation, backtests, and stakeholder-ready handoffs.

Governance & Monitoring

• Model validation: train/validation/OOT, calibration checks, score stability, drift monitoring, and portfolio KPI readouts.
• Operational controls: observability, eval/tracing, least-privilege access, and workflows that business, risk, and operations teams can inspect.

More on charlieviettq.github.io — essays and tutorials on credit risk modeling, decisioning, and production AI.

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