Johan Vaz (JVAZ) — Initiating Coverage

Investment thesis§ 1

Ships whole systems, not snippets. Six of the projects below are live in production right now — ingestion, API, frontend, deploys. You can break them yourself; the links are right there.

Treats trust as an engineering problem. Chunk-level citations, refusal behavior tested like a feature, eval suites written before the demo. AI that sounds right is cheap; AI that can show its receipts is the scarce asset.

Reads the 10-K and writes the parser. The tools mirror how analysts actually work — risk decomposition, regime context, covenant headroom — because the builder studies the finance, not just the framework.

Coverage universe§ 2

System	Status	What it proves
Equity Research Copilotsource ↗ · cited RAG · keyless demo	● LIVE	Document-grounded research over SEC filings with chunk-level citations, memos, and comparisons. The public demo runs keyless on deterministic cited synthesis — ask it about NVDA's data center revenue.
Market Regime Dashboardsource ↗ · FastAPI + React	● LIVE	Macro and risk context with transparent regime rules and freshness checks — the same API powering the live strip at the top of this note.
Portfolio Risk Copilotsource ↗ · VaR / ES / stress	● LIVE	API-first risk engine: VaR, expected shortfall, variance contributions, concentration flags, stress scenarios, plain-English commentary. Try it below without leaving this page.
finance-labssource ↗ · 10 diagnostics · CI green	● LIVE	Ten offline, test-covered risk diagnostics — margin cascades, ETF liquidity stress, option skew, covenant headroom, central-bank path — with a gallery of real CLI output.
Financial LLM Eval Harnesssource ↗ · 50 cases · published report	● LIVE	Evaluation suite for financial QA systems — with a published report of it evaluating the live copilot demo above, including the refusal bug and coverage gap it caught. Failing honestly, in public.
Curiovoice · reasoning agents	CODE	Learning-by-teaching, instrumented: teach an AI novice by voice, agents map your claims against a curriculum, the novice teaches it back with your gaps exposed.
FluentAIevaluator/memory agents	CODE	Agentic language tutor: adaptive lessons, real-time speech, spaced repetition driven by actual conversation mistakes. 30+ iterations of visible development.

Exhibit A — run the risk engine from this page

POST portfolio-risk-copilot-pi.vercel.app/analyze

This isn't a screenshot. The button below sends a portfolio to the subject's deployed risk API and renders whatever it says — including when it disapproves of the concentration.

Catalysts§ 3

Near-term: one flagship system built in public over eight weeks — current candidates are an event-driven backtesting engine with an automated look-ahead-bias test suite, and an MCP risk server exposing the risk engine to any AI agent. Ongoing: new diagnostics land in finance-labs (ten and counting), and every deployment on this page auto-ships from git push.

Risk factors§ 4

Subject is a student; production experience is self-inflicted rather than employer-supervised. Mitigant: the systems are, in fact, in production.

Backtests and demos on this page are research artifacts, not investment advice; no returns are claimed anywhere. The analyst considers this a feature.

Coverage universe is early-stage: histories are months, not years. Mitigant: velocity — the gallery, monorepo, and this note itself shipped inside a single week.

Enthusiasm for automation may result in your workflows getting quietly improved. No mitigant available.

Analyst certification

The analyst certifies that the views expressed in this note accurately reflect his own capabilities, that every system marked LIVE is genuinely deployed and was verified working before publication, and that no language model was permitted to invent a number on this page — the live ones come from the APIs.