We let the AI run a portfolio. In public.
A team of Claude agents, three analysts, a portfolio manager and a macro strategist, manages a ₹10 lakh paper portfolio of NIFTY-500 stocks on our own capital. Every evening it reads the filings and the news, decides, and writes down why. Every return below is time-weighted with costs on. We publish it on the days it loses too.
Paper-traded, no live orders yet. Not investment advice. Not open to outside money.
The portfolio is behind the index so far. We show that because a scoreboard you can edit is not a scoreboard. The bar we set, three to five points a year over the NIFTY 500 across two market regimes, is not yet met, and four months is far too short to judge either way. Full scorecard on the dashboard.
A research tool should be able to
put its money where its numbers are.
Can a disciplined team of AI analysts, reading the same filings and news our terminal serves, beat the index net of realistic costs? Not in a backtest. Day by day, forward, with the rules fixed in advance.
The terminal promises every number is proven. The lab is where that promise gets expensive. If the research is wrong, the portfolio says so in rupees, in public.
Which signals carry information and which are noise. The scorecard tracks each analyst's hit rate at 5 and 10 days, so we promote or drop inputs on evidence, never on a hunch.
Seven steps, every
trading day, after the close.
The models do the reading and the judgement. Code does the gating, the risk limits, the tax maths and the fills. That boundary is deliberate: an AI may decide for our own book, but it never gets to bend a rule.
- Macro strategist · one brief per session, prepended to every prompt
- Devil's advocate (Claude Haiku) · argues against each decision, in shadow mode
- Holdings correlation and F&O-ban tags · flagged, not yet enforced
Anything that could change a decision ships in shadow first and is promoted only after at least 30 trading days of scorecard evidence.
-
1
Gatedeterministic
Of ~150 tracked stocks, only those with fresh filings, news or a price event (typically 11 to 66) go to the analysts. The rest hold. Attention is rationed, like a real desk.
-
2
Fundamental analystClaude Sonnet
Reads the sector card, the filings pack, valuation versus peers and analyst consensus. Returns a thesis with growth and risk grades and a conviction score.
-
3
Sentiment analystClaude Sonnet
Reads event-classified headlines, management tone and the earnings surprise. Returns tone, drivers and conviction.
-
4
Technical analystClaude Sonnet
Reads indicators and derived pattern labels. Returns signal, setup and conviction. Price context only, never the thesis.
-
5
Portfolio managerClaude Opus
Synthesises the three views against current positions and cash, inside a macro brief the strategist writes once per session. Proposes buy, trim, sell or hold with size.
-
6
Risk and tax overlaydeterministic
Code, not a model, applies the caps: 5% per stock, 25% per sector, the cash rule, three new positions a day. Then estimates the tax cost of any sale. The model cannot override it.
-
7
Fill and publishdeterministic
Approved orders fill at the next open with transaction costs and slippage. The decision, the reasoning and the trace are pushed to the public log and the dashboard.
Fixed in advance. Enforced in code.
A track record is only as honest as its accounting. These are the constraints the portfolio has lived under since day one, and the bar it has to clear before we call it anything.
More filings than any analyst
reads in a career.
Every decision carries the analyst views, the manager's reasoning, the risk verdict and the tax note. Read today's log.
- A research experiment on the founder's own capital.
- A forward, out-of-sample test of the same data the terminal serves.
- Fully logged: every prompt, view, decision and fill is versioned.
- Investment advice, a tip service or a model portfolio to copy.
- Open to anyone else's money. We are not SEBI-registered and do not manage client funds.
- Live yet. Fills are simulated at the next open with real costs; no broker is connected.
- 1Now: paper trading, scorecard accumulating, shadow agents under evaluation.
- 2Next: live execution of the same decisions, on our own capital, through an Indian broker.
- 3Then: the full decision log and scorecard published here, so the terminal's research can be judged by what it did, not what it said.
The analysts run on Claude Sonnet, the portfolio manager on Claude Opus, the devil's advocate and news classifier on Claude Haiku. The gate, the risk limits, the tax maths and the fills are plain code.