Evidence-first beats answer-first.
In any regulated domain, make the model point at sources or stay silent. Ask every vendor for the evidence chain, not the answer.
The Lab · Case Study 01
The research engine that told me I was losing.
Most AI in investment research is a demo wearing a suit. It hallucinates a thesis, flatters your positions, and calls it insight. I wanted the opposite: a system whose entire job is to make the evidence boring, checkable, and impossible to spin.
I set the constraints and the acceptance bar. AI agents did the building. I audit the results. The first thing the finished system did was tell me my own dashboard had been lying to me.
Running dailyPaper portfolioResearch only, not advice
The business case: AI that flatters is AI that fails
Every regulated domain has the same problem with AI. The technology is optimized to please the person asking, and the person asking usually has money on the line. That combination produces confident nonsense with your logo on it.
In investment research I see three failure modes repeat. Answer-first AI that gives you a verdict and hides the evidence, so nobody can check anything. Math without a trail, dashboards showing outcomes no auditor could reconstruct. And cost sprawl, AI systems that quietly burn budget because nobody set a ceiling at design time.
FundHouse is my lab's answer to all three, built for the Indian equity market: an AI stock research engine where every answer points at its evidence, every number shows its work, and the whole system stays inside free-tier cost ceilings by design.
Nobody checks math that makes them happy. So the system has to check for you.
The day my dashboard lied
The portfolio page said I was up three lakh rupees. The truth, once the money-weighted return was computed properly: I was down 6.55% while the benchmark was down 7.07%. The difference wasn't a rounding error. The interface had been counting my own monthly deposits as gains.
The uncomfortable part is that the wrong number felt true. It was positive, I had worked hard, the market was down. Of course I was beating it. The system had simply agreed with me, which is the most dangerous feature an AI can have.
So the product got rebuilt around one rule: every figure on screen opens into its evidence. Dated flows, fills, prices, sources, all visible in the interface. Honest XIRR against the NIFTYBEES benchmark, drawdown on the holdings series, and an audit trail a CFO could love. The math is now harder to admire and impossible to spin.
How the engine works
Every morning before the market opens, a scheduled pipeline builds a fresh evidence base across 250 Nifty-listed companies: fundamentals, governance, shareholding, technicals, news, and web intelligence, assembled into a per-stock data pack.
A weighted multi-factor model scores the universe. Resilience and quality carry 35%, growth 25%, industry tailwind 15%, governance 10%, valuation 10%, liquidity 5%. The weights are explicit, so a ranking is never a mystery. It's an argument you can interrogate.
A paper portfolio executes on that research with real discipline: fills, ledger, cash flows, money-weighted XIRR against NIFTYBEES. A chat interface answers questions like "why is this stock ranked here?" by pointing at the evidence base, grounding first, opinion never. If the evidence doesn't support an answer, the guardrails make it say so.
The architecture, in the open
A cloud-first pipeline engineered to run on free tiers. The constraints are the design.
- Universe
- 250 Nifty-listed companies refreshed on six freshness tiers, from 15-minute indices to yearly filings
- Evidence packs
- Per-stock packs covering identity, market, fundamentals, governance, shareholding, web intelligence, and technicals, with 15 indicators including RSI-14, MACD, Bollinger, ATR, Ichimoku, and VWAP
- Scoring
- Weighted multi-factor model with explicit weights. Rankings are inspectable arguments, not vibes
- Honest books
- Paper portfolio with fills and ledger, money-weighted XIRR vs NIFTYBEES, drawdown on the holdings series, and every figure opening into its evidence table
- Grounded chat
- Answers point at the evidence base or decline to answer. Guardrail checks and a golden-set evaluation suite replay on every change
- Infrastructure
- Next.js on Vercel, Neon Postgres, Cloudflare Workers dispatcher, GitHub Actions, LiteLLM gateway, self-hosted search, FinBERT sentiment
- Cost discipline
- Built to free-tier ceilings. The system wakes on schedule, does its work, and sleeps. Single-digit monthly run cost, a discipline that maps directly onto enterprise unit economics
What this proves for client programs
The lab exists so these patterns arrive in enterprise conversations with receipts attached.
Honest math is a product feature.
Money-weighted returns with dated flows, drawdown on the real series, numbers that open into their workings. If a metric can't show its work, it's decoration.
Evals are the real CI.
LLM features get golden-set suites that replay on every change, treated like regression tests. Confidence in AI is engineered, not asserted.
Cost ceilings by design.
Free-tier-first architecture forces a discipline that maps one-to-one onto enterprise unit economics. The cheapest system sleeps when it's done.
Scope and discipline
- Research and education only
- Paper portfolio, no real-money execution
- Broker execution disabled at the dispatcher level, by design
- No performance claims, by policy
- Auth-gated, running privately
The bridge to your program
Running AI where the rules are unforgiving?
Regulated domains punish answer-first AI. I help teams design evidence-first systems, with governance, evaluation, and honest metrics, that survive both auditors and users. If that's your problem, it's the conversation I enjoy most.