Hi, I'm Sang.

Built at Citi to catch and escalate payment control failures across 52 global applications — processing ~25M records monthly and validating roughly $1.5 trillion in payment value every cycle. Designed and built from scratch as the sole developer: a config-driven Python platform running 4 pipeline types — Threshold Validation, Duplicate Detection, Dual Blind Rekey, and GPOC executive reporting — each with its own business rules per application. Reduced ~90% of manual effort per monthly cycle, saving 150+ hours. The codebase is internal to Citi — the architecture and outcomes are mine to discuss.

Internal tool — codebase confidential per Citi policy

A retrieval-augmented extraction service that reads a 200+ page commercial credit agreement and returns the key terms — borrower, facility amount, maturity, pricing, covenants, events of default — as structured data. The hard part isn’t getting an answer from a language model; it’s knowing whether to trust it. Every value comes back with the page and the exact sentence it was taken from, and each citation is verified in code against the retrieved text, so an unsupported answer is downgraded to “not found” rather than invented. Search runs in two passes — vector similarity over pgvector/HNSW, narrowed by a cross-encoder re-ranker running locally — and ingestion is queued through Kafka with an idempotent worker, so uploading the same document twice is safe. Accuracy is measured, not assumed: an evaluation harness scores results against hand-labeled answer keys, and was used to lift accuracy on a real SEC filing from 57% to 71% — and to disprove a change that turned out to make no difference at all.

A live financial data monitoring service built and deployed on AWS — ingesting U.S. economic indicators daily via the FRED API, detecting anomalies, and serving alerts through a secured FastAPI layer backed by PostgreSQL. Built specifically to close real gaps: AWS infrastructure provisioned with Terraform so the entire server can be rebuilt from scratch in under 2 minutes, GitHub Actions CI/CD so every push runs tests before touching production, and pytest coverage maintained across 3 months of active development. Completed a pre-launch security audit — 6 vulnerabilities identified and resolved before the first public deployment.

A fully automated ELT pipeline that ingests daily stock prices from Yahoo Finance, transforms them through a medallion architecture using dbt, orchestrates every step with Apache Airflow, and serves a live Metabase dashboard — all running inside Docker. Every weekday at 6 PM the pipeline triggers automatically: raw prices land in PostgreSQL, dbt rebuilds the star schema across 3 data layers (Bronze → Silver → Gold), 26 data quality tests validate the output, and the dashboard updates itself. Built to demonstrate the full data engineering lifecycle from ingestion to visualization.

Built from scratch as a first-time parent who couldn't find one app that did everything. ThriveKid is a deployed full-stack PWA — 12 feature modules, JWT authentication with refresh token rotation, a timezone-aware recurring event engine, and WHO percentile growth charts. Clean Architecture across a 3-layer .NET 8 backend, React 19 TypeScript frontend deployed on Vercel, Docker multi-stage builds, and GitHub Actions CI. Went through two full mentor code reviews with a senior engineer — 25+ issues identified, all resolved, including a 13-issue security audit conducted before public launch. The project that taught me what good engineering actually looks like.