I build AI systems for financial research and risk — the unglamorous machinery that decides whether an AI tool is trustworthy or just fluent: eval harnesses, cited retrieval over SEC filings, and risk analytics that explain themselves.
Computer Science + Quantitative Finance at NUS. Based in Singapore.
Site: johan-vaz-site.vercel.app — several projects below are deployed and clickable, not just readable.
Currently: looking for internships or junior roles in AI agents, LLM evaluation, fintech, or quant research tooling — ideally on a small team that ships.
| Project | What it is |
|---|---|
| AI Equity Research Copilot | Document-grounded research over SEC filings: EDGAR ingestion, chunk-level cited retrieval, structured memos — plus a finance QA eval set so citation precision, refusals, and hallucination risk are measured, not assumed. |
| Financial LLM Eval Harness | 50-case evaluation suite for financial QA systems: factual extraction, multi-document synthesis, refusal behavior, adversarial prompts, scoring, and regression reports. |
| Curio | Learning-by-teaching, instrumented: teach an AI novice by voice, reasoning agents map your claims against a curriculum, then the novice teaches it back using only what it learned from you. |
| FluentAI | Agentic language tutor with adaptive lessons, evaluator/memory agents, real-time speech, and spaced repetition driven by actual conversation mistakes. |
| finance-labs | A toolkit of small, offline, test-covered risk and market-structure diagnostics: margin cascades, ETF liquidity stress, option skew, factor crowding, covenant headroom, and more. Live results gallery → |
| US Market Regime Dashboard | Macro/risk context dashboard with transparent regime rules, data-freshness checks, and exports. FastAPI + React. Live → |
| Portfolio Risk Copilot | API-first portfolio risk: VaR, expected shortfall, correlations, concentration flags, stress tests, plain-English commentary. Live → |
- Grounded, or it doesn't ship. Generated claims trace to source chunks; when evidence is weak, the system refuses — and that behavior is tested like a feature.
- Evaluated, not vibe-checked. I build the eval harness before trusting the output: golden sets, regression scoring, adversarial cases.
- Honest about failure modes. Finance punishes overconfidence, so I document where tools break and what they must not do.
- Email: v.johan2234@gmail.com
- LinkedIn: sg.linkedin.com/in/johan-vaz