I build evaluated AI systems and financial research tools (cited retrieval over SEC filings, financial QA evals, inspectable risk analytics), and I contribute to open-source tools, with merged work in T3 Code and Apache Arrow.
NUS Computer Science with a Quantitative Finance minor; graduating May 2029. Based in Singapore. Authorized to work in Singapore. Looking for a 2027 internship, remote preferred, especially with small teams that build with coding agents: either 15–20 h/week Jan–Jun 2027 alongside classes, or full-time May–July 2027.
Site: johan-vaz-site.vercel.app
- T3 Code — 10 merged fixes to an AI coding-agent app. Reliability and usability fixes across the web client, CLI, provider and source-control integrations. Examples: opening files from Claude and ACP tool-call diffs, restoring long-path file suggestions on web and mobile, and preserving empty replacements in OpenCode edits.
- Apache Arrow — numerical CSV imports. Fixed
+1and+1.5being inferred as text; added regression coverage for valid numbers, overflow, and malformed signs. Merged into the Rust implementation on 17 September 2026. - Apache Arrow — fragmented-mask selection. Added adaptive dispatch to interleave kernels, preserving null handling and the contiguous-mask path. The PR's 8,192-row i32 benchmark with random 50%-true masks recorded 76.889 µs → 33.676 µs: 56.2% less execution time for that workload.
- bt — integer allocation with nonlinear commissions. Fixed a search that skipped an affordable share quantity and raised an error; added a regression for minimum, per-share, and capped commissions.
- ffn — multi-period annualization. Corrected frequency inference that inflated annualized ratios for 30-minute data, with Sharpe regression coverage.
I use AI coding tools. The linked PRs contain the implementation, review discussion, validation, and AI-assistance disclosures where applicable. Benchmark results apply to the recorded workloads and environment.
| Project | What it is |
|---|---|
| AI Equity Research Copilot | SEC EDGAR ingestion, chunk-level cited retrieval, and structured research memos; a synthetic finance QA suite checks citations, refusals, and unsupported claims. |
| Financial LLM Eval Harness | 50-case financial QA suite covering factual extraction, synthesis, refusals, and adversarial prompts, with scoring and regression reports. |
| Curio | Teach an AI novice by voice or text; curriculum-grounded claim checks, spoken teach-back, and an inspectable learning report. |
| FluentAI | Language-learning demo with adaptive lessons, conversation, persistent learner memory, and spaced review. |
| finance-labs | Offline Python risk and market-structure diagnostics using transparent synthetic scenarios. Results gallery. |
| US Market Regime Dashboard | Macro/risk dashboard with documented regime rules, source-freshness checks, and exports. Live dashboard. |
- Email: v.johan2234@gmail.com
- LinkedIn: sg.linkedin.com/in/johan-vaz