Agentic Operating Systems for Dealmakers
Manthan Intelligence builds agent operating systems for sell-side advisory and buy-side investing. Narada surfaces the deal months before it hits the wires. Twelve analytical lenses pressure-test it independently. The memo arrives banker-grade. And every deal, every signal, every miss makes the system sharper — judgement that compounds in the firm instead of walking out the door.
Before anyone logged in
Nobody prompted anything. These are the scheduled, autonomous firings of the last 24 hours — function labels only, the same anonymised feed that drives the live proof page. Times are UTC; the marker is now.
source: firings feed · as of 2 Jul · auto-generated, never hand-edited
The Problem
A mid-market banker starts the year with a hundred opportunities and can deeply track twenty. A pod of twelve runs three deals at a time. And the generic AI everyone bought gives one opinion, with no memory and no method.
One hundred names in the book; bandwidth for twenty. The other eighty get covered late, shallow, or not at all — and the mandate goes to whoever arrived early with the sharper idea. Coverage is the constraint, not judgement.
Most of a junior analyst's week disappears into Crunchbase, PitchBook, LinkedIn and filings — hunting signals a machine should surface. The expensive hours go to data collection, not advice.
A single agent produces a single coherent narrative. Coherence isn't accuracy. Real committees work through structured tension between mandates — generic agents have none.
Deal teams, diligence committees and advisory panels succeed through structured disagreement between different mandates. We made that process synthetic, autonomous — and measurable.
Don't take the claim — watch it
A compressed replay of the analytical flow on an anonymised composite deal — three of the lenses shown. Note what a single-agent answer would have hidden: the disagreement.
Input: 18-page deck — B2B logistics marketplace, Series A, growth-stage raise, claims multiple-x YoY GMV growth.
Press Run the council to replay the flow: evidence base → independent lenses → synthesis.
GMV trajectory verified against the deck's own cohort tables deck p.7; well above the sector median for this stage KG comparables.
High alignmentContribution margin negative at the current take rate deck p.12, recomputed; CAC payback unproven — the cohort table stops where it gets interesting.
Partial alignmentBoth sides multi-home; most supply also lists on the two incumbents KG relationship scan. Scale without lock-in is a subsidy programme.
Low alignmentThree lenses, three honest answers. The synthesis layer doesn't vote — it locates the tension that decides the case:
Composite illustration; companies anonymised; figures are illustrative, not production metrics. The production flow runs twelve lenses plus extended synthesis. SEBI-safe alignment language throughout — never an instruction to act.
Products
The same multi-agent core — evidence, deliberation, synthesis, memory — pointed at the two sides of every deal. Every mandate it touches makes it sharper.
Sell-side · M&A and fundraising advisory
An entire deal team, one banker. Narada, the origination agent, watches your coverage universe 24/7 and flags transactions three to six months before they surface — with warm relationship paths and a banker-grade positioning memo. Behind it, independent analytical lenses pressure-test every thesis before it reaches a client.
For: boutique and mid-market M&A bankers, fundraising advisers, solo MDs
Buy-side · funds, family offices, syndicates
The system already runs the analytical flow of an active venture operation — 49 agents across 8 divisions, screening deals, drafting memos through twelve independent lenses, scoring its own calls against real outcomes every night. Not a demo. A working buy-side pod you can watch in real time.
For: VCs, family offices, fund-of-funds, angel syndicates
The Methodology
The architecture of the best decision-making bodies — running autonomously, on your mandate.
01 — EVIDENCE
Documents, data and context exhaustively extracted into a structured base all analysis builds on. Every material fact, every assumption, every gap.
02 — DELIBERATION
Multiple agents with genuinely different mandates evaluate the same evidence. Growth vs. risk. Quantitative vs. qualitative. Real tension, not theatre.
03 — SYNTHESIS
A synthesis layer finds the tensions that matter and preserves disagreement as signal. You get a decision map, not a vote count.
04 — MEMORY
Every analysis enriches a knowledge graph of 110,795 entities. Next month's work is better than this month's — and the learning lives in the system, not in someone's notebook. People can leave; the judgement stays.
Start Here
The Narada Executive Summary and the 1000x Banker brief live in the resource library — sign in with Google, LinkedIn, or a one-time email code and they're yours. No password to invent, no card, no calls until you ask for one.
Narada — Executive Summary 2 pages
The 1000x Banker — Product Brief overview
Open the resource library →Free account · one-time code by email · 30 seconds
Pricing Context
A VP, all-in
~£400K/yr
PitchBook
$12K–70K/seat/yr
Harvey (Legal AI)
$12,000/seat/yr
Manthan Intelligence
A fraction of one VP
Three extra mandates a year is £4.5M+ in new fees from one VP's coverage. The hunting is what we automate — the judgement stays yours.
Data Security
IP leakage is the #1 concern with AI adoption in advisory work — mandates are confidential by definition. The architecture eliminates it by design: three tiers of memory with hard boundaries. Confidential context never crosses organisational walls.
Processed, never trained on. Analysis runs on enterprise-tier AI under commercial terms: your data is used to produce your output, is excluded from model training, and never improves anyone else's system.
BYOAPI option: plug in your own Anthropic API key. Exploration runs on your infrastructure, your billing, your data governance.
Mandate details, cap tables, founder disclosures — visible only to the specific engagement. Never crosses to other clients or the shared graph.
Your firm's accumulated intelligence — coverage assessments, decision history, relationship context. Private to your firm, full stop.
Sector patterns from public sources. No company-specific data, no attribution. The shared layer that makes every analysis richer.
Self-Calibration
Most AI products ship and hope. Ours runs a daily blind assessment against real outcomes, sweeps calibration weekly, and records every miss as a structured learning entry. The numbers below update from the production system — including the ones that aren't flattering.
In production with an active venture operation, a fundraising advisory firm, and a forming cohort of sell-side design partners in London and New York.
Early Users
“Every deal that reaches our investment partner now arrives with twelve independent reads, the comparables, and a draft memo — work that would take an analyst a week or longer is on the table before our morning call.”
“The target company screening pack — market context, investor fit, the questions to ask — lands in hours, not days. And the early origination signals point me at conversations I would have found months too late.”
“Really very good and accurate — it saved me six to ten hours, and it needed almost no rework.”
Charaka Notes
Intelligence dispatches from the running knowledge graph. Five pillars, five days a week — patterns surfaced by structured analysis of thousands of companies and hundreds of postmortems. Plus the founding note on why we built this.
Read the latest →Free. 5 dispatches/week. Unsubscribe anytime.
A briefing is a working session, not a sales call: bring your mandate focus and we'll show you what the system would have flagged this quarter.
Prefer to read first? The two-pager is in the resource library.
Received.
We'll come back to you within one business day. If it's time-critical, write to [email protected].