Sang Thai.

A meeting assistant that listens while the meeting is still happening and answers in the moment. Speech streams through the Deepgram API with speaker labels; the Anthropic API turns it into live answers, recaps, and answers across past meetings that name the meetings they came from.

I launched it on my own, with paid plans. Margin is protected in code: each plan has a usage allowance and a daily spend breaker caps total model cost, and paid access is granted only by a signature-verified Stripe webhook. The backend is FastAPI on Render, the frontend React and TypeScript on Vercel, and the data lives in Supabase Postgres with row-level security. pytest and Playwright end-to-end suites run in CI, which gates every backend deploy.

I designed and built ATLAS end to end, a production-grade Python platform for a 50+ application payment-controls program at Citi. It automates approval-threshold, duplicate-payment and dual-blind-rekey control tests, each with its own business rules per application. Shared modules ingest, validate and export Excel, CSV and JSON files, and each new application needs only a JSON config entry plus template scripts. Monthly testing went from about 100 hours to under 5, and full-population tests now cover about $1.5T in monthly payment value on top of the manual monthly samples. After a reorg, ATLAS is now owned by another team. The code is internal to Citi; the architecture and the outcomes are mine to discuss.

Internal tool — codebase confidential per Citi policy

A question-answering workspace for a team’s own files. Upload documents and spreadsheets, ask in plain English, and every answer comes back one of two ways: quoted, with the source documents listed underneath it, or computed — Avocado writes Python, runs it against your data in a locked-down container, and shows you the code behind the number.

It is careful about what it does not know. Asked what policies a workspace defines, it lists the three it found, flags that two of them are titles with nothing behind them, and closes with what is missing — the actual rules, effective dates and scope are not in these documents. Ask for an executive summary and it charts every spreadsheet in the workspace, then ends the same way.

Capabilities are switches rather than magic: each integration shows what it will cost in context before anyone turns it on. The sandbox is the part I would defend in review. Generated code runs with no network, a read-only filesystem, and hard memory and CPU caps — applied to every run, never set per request, because a caller that can choose its own limits can choose none, and 16 container-escape tests attack it to prove it. Role-based access and tenant isolation are enforced in code, checked by 9 cross-tenant tests, so no team can read another’s data. Tools are called over MCP, and a model router works across LLM providers (Anthropic, OpenAI, Ollama).

A service that reads a 200-page commercial credit agreement and returns the terms a credit team actually needs — borrower, facility amount, maturity, pricing, covenants, events of default — as structured data. Every value arrives cited, with the page and the exact sentence it came from. Code then checks that every citation points at a passage that was actually retrieved, so anything that cannot be traced comes back not found rather than invented.

Search runs in two passes: vector similarity over pgvector to gather candidates, then a cross-encoder re-ranker running locally to narrow them. 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 answers against hand-labeled keys — on a real SEC filing, accuracy measured 71%, up from a 57% baseline, and it also disproved a change I expected to help, which turned out to make no difference at all.

A statistical comparison of two anti-money-laundering detectors — a rule-based baseline and a graph-based method — on IBM’s synthetic AML dataset (HI-Small): 5 million simulated transactions across 515,000 accounts. It is a research benchmark on synthetic data, not production AML work. The question is not which number is bigger but whether the difference is real, so every metric comes with a bootstrap confidence interval, AUC is compared with DeLong tests, and p-values are corrected across metrics.

The honest answer is unflattering to both. At a matched flag rate the graph method catches 48% of laundering against 29% for the rules, yet both flag hundreds of innocent transactions for every real one. The stack is built like production: PostgreSQL with least-privilege roles, a job queue and worker, a FastAPI API, a React and D3 network view, non-root Docker images, and CI with pip-audit, Bandit and 209 tests.

A monitoring service that tracks three U.S. economic indicators — the federal funds rate, CPI and unemployment — pulling them daily from the FRED API, storing them in PostgreSQL, and raising an alert when a value moves further than its configured threshold.

It is built the way production services should be. The whole environment is defined in Terraform — EC2 instance, firewall rules, key pair — so it stands up from nothing with one command instead of being a machine nobody dares touch. Deploys are gated on the test suite: the deploy job declares a dependency on the tests, so a failing push cannot reach production. And I ran a security audit before the first public deployment, which found six issues I fixed before anyone else could see them.

It ran on AWS EC2 for about five months, until August 2026. I took it down deliberately: the infrastructure code is the artefact worth keeping, and leaving a t2.micro running to serve a demo nobody is querying is a bill, not a portfolio piece. terraform apply brings it back.

An ELT pipeline that turns daily stock prices into a dashboard without anyone touching it. Every weekday at 6 PM Airflow starts the run: prices land in PostgreSQL from Yahoo Finance, dbt rebuilds a star schema across three layers (Bronze → Silver → Gold), 26 data quality tests check the output, and the Metabase dashboard picks up the result. The whole stack runs in Docker.

It follows the medallion pattern rather than the shortest path from source to chart, and SQL window functions with ROW_NUMBER() de-duplication make every rerun idempotent.

A full-stack PWA that helps parents track and support a child’s development: 12 feature modules, JWT authentication with refresh token rotation, a timezone-aware recurring event engine, and WHO percentile growth charts. I built it as a first-time parent who could not find one app that did all of it.

Clean Architecture across a 3-layer .NET 8 backend, a React 19 TypeScript frontend on Vercel, Docker multi-stage builds and GitHub Actions CI. It went through two full code reviews with a senior engineer — 25+ issues raised and resolved, including a 13-issue security audit before launch.