An AI research desk for Indian equities.
A team of ten AI agents runs a disciplined, catalyst-driven process over Indian listed equities: sourcing ideas, stress-testing them through nine gates, and proposing positions. Built entirely on public information. It never trades; the system proposes and a human makes every decision.
A run, live.
What a research run looks like from the inside. Names and figures are illustrative placeholders by design; the real logs stay private.
Discipline is the edge.
Catalyst investing is not a novel idea. The hard part is doing it the same way every time, without improvisation or emotion.
The strategy buys companies where a specific, datable, public catalyst is likely to drive a step-change in future revenue or earnings that the share price has not yet reflected. The entire thesis rests on one question: is there a large, confirmed, forward-looking change the market has not fully priced in?
People are bad at running that question repeatably. We anchor, we hope, we hold losers and sell winners. So the process is encoded as software: a fixed catalyst definition, a hard quality gate, a timing gate, strict sizing, and pre-committed exits, executed identically on every idea. The machine supplies the discipline; the human supplies the judgement and the final call.
Ten agents, one desk.
Nine independent analytical agents in sequence, each with its own context, each scoring and voting. Watch the desk work below: the live badge is the gate currently judging a candidate.
Origination
Scans public sources for a real, not-fully-priced-in catalyst.
Sector
Does the sector cycle and backdrop support the idea?
Market
Can this company actually win and execute the catalyst?
Fundamental
Financials, governance and a 14-sector KPI playbook.
Technical
Timing gate: base, breakout, volume and delivery. Cash only.
News-confirm
Is the catalyst live, authentic, material, and not already spent?
Risk
Break-the-trade review; sets the risk envelope and stops.
Portfolio
Sizes the position within the risk ceiling; targets and phasing.
Principal
Final go / no-go. Reviews the whole chain and issues the order ticket.
Investor Relations: the human-in-the-loop
Publishes a one-pager for every idea and carries the operator's feedback back in, so a killed name can be resurrected on new evidence. It reports; it does not gate.
From signal to decision.
A run loads the operating context, ingests candidates, pushes each through the gauntlet, and ends at a Principal decision with a full score log either way.
Source
Overnight filings, order wins, tenders, policy, press, or a manually fed name.
Gauntlet
The nine analytical agents each score and vote, independently.
Decide
The Principal weighs the full chain: go, watchlist, or kill.
Ticket
On a go: an execution-ready order ticket with size, entry, target, stop and reasoning.
Log & learn
Everything is logged; the operator's feedback re-runs candidates next time.
Four ways a stock re-rates.
Every idea must carry at least one public, datable catalyst, and survive the master filter. Most don't. Watch the funnel below.
Order-book re-rating
A confirmed order backlog that's large relative to today's revenue, implying future revenue the market hasn't capitalised. Signed orders only.
Government tender wins
A large public-sector award (including foreign governments) that re-rates the company on contracted, visible revenue.
Synergistic M&A
Friendly deals where both sides re-rate, read only from public filings and disclosure thresholds, never private information.
Policy tailwinds
Announced policy that structurally lifts a sector (incentive schemes, mandates) and the listed companies positioned to capture it.
The stack.
Claude, as a team of agents
Each agent is a written charter with a job, inputs, a structured output and a confidence format: ten specialists rather than one prompt.
An Obsidian vault
The thesis, agent charters, sector KPI playbooks and a running decision log: the system's long-term memory and rulebook.
Airtable
Candidates, per-agent outputs, positions, order tickets and a calibration log: the structured record of every run.
Public sources only
Exchange filings and disclosures, screeners, and news/sentiment feeds. No private or unpublished price-sensitive information, ever.
Rules that don't bend.
The design, not a track record. Sizing and exits are pre-committed so no single idea can do outsized damage.
Conviction-weighted, descending ladder. One concentrated swing is allowed; no name can dominate the book.
Hard and pre-committed. A broken thesis exits even before the stop hits.
A drawdown from peak halts new entries and escalates to the human for a decision.
Short-to-medium holds, then rotate the capital into the next opportunity.
Booked and recycled, rather than held for the long term.
A position stays within ~1–2% of a stock's 20-day average traded value.
Judged on risk-adjusted return vs the Nifty Smallcap 250, where most of the universe lives.
What it adds up to.
Nine analytical gates plus a human-in-the-loop wrapper.
Order book, tenders, friendly M&A and policy, plus a QARP lane.
Curated metrics the fundamental agent scores each name against.
Every signal is built on public disclosure. The system proposes; the human executes.
A method, not a promise.
This is a strategy and research system, not investment advice and not a solicitation. Markets carry real risk of loss. Nothing here is a performance claim: the interesting work is the process, encoding a repeatable, disciplined, public-information-only method with a human firmly in the decision seat.