Rollupa buys and keeps fire safety and compliance-service firms from retiring owners. An agent team finds each firm, underwrites it, argues with itself and presents a decision a committee can act on. Software is how we get there faster than anyone else.
The owners are directors in their sixties running firms with £500k to £3m of revenue. Their options are to wind down, sell to a local competitor at a poor price, sell to staff who rarely have the cash, or carry on past the point they want to.
Private equity cannot help. Below about £1m of EBITDA the deal costs and management time do not work. Capital partners want exposure to these cash-generative businesses and cannot source, underwrite or integrate them at this size.
Python owns the stages, retries, state and resume. Agents work inside each stage and write only through typed deal-file tools. Every claim carries a source or a derivation. Every sum comes from a finance function, never from a model.
Companies House is searched for active firms trading 15 years or more, then filtered to those where every director is 60 or over, and sized by proxies.
A deal lead plans and dispatches an analyst, an AI operations specialist and a structurer. The lead raises challenges when their numbers conflict.
About 25 deterministic checks, then an agent that re-fetches sampled sources and scores a rubric. Failures go back for a fix loop.
Three personas with different incentives vote blind, debate and send deals back. Approval means an indicative, non-binding offer subject to diligence.
A ranked pipeline, a memo per decided deal, a draft seller letter marked DRAFT, NOT SENT, and a run report with cost and retries.
Every stage is checkpointed and idempotent. In the recorded run we killed the process mid-flight and it resumed from state.
The recorded run covers Greater London on live Companies House data. It was started from nothing and ran to completion.
Of 10 deals taken to committee, 1 was approved and 7 rejected, 2 failed after fix loops. Figures come straight from the run report. The pipeline was killed mid-run once and resumed from state.
The revenue backtest is the weak point and we report it as such. The headcount proxy undershoots because firms that disclose turnover skew larger than our target band. Medium-confidence estimates, which rest on headcount alone, were inside range only 10% of the time. The fix is to recalibrate on firms we own, which is why completions feed the estimator.
Venture money funds the engine and never buys a business. Sellers, lenders and private equity fund the acquisitions, so each layer of capital sits where it is cheapest.
| Layer | Funded by | Pays for |
|---|---|---|
| Platform equity | Venture, seed and Series A | Engine, central team, shared back office |
| Seller financing | Retiring owners | A large slice of each price through loan notes and earn-outs |
| Senior debt | Private credit and banks | Part of each price, secured on cash flows |
| Deal equity | PE, family offices, angels | The remainder of each price |
| Selective exit | PE buyers | Vertical clusters, only when the price beats holding |
The end state is a permanent compounder in the mould of Halma or Constellation Software. Illustrative economics: a £1.5m revenue firm bought at 4x £225k EBITDA for £0.9m, lifted to a 25% margin by the shared back office, is worth about £3.75m inside a platform valued at 10x. These are assumptions to test, not facts. The deck sets out the moat, the scale check and the risks.
This site shows masked output only. Firm names, addresses, company numbers and websites from the run are not published.
Background in AI and machine learning, which builds the engine, aerospace, which brings a safety-critical and compliance-heavy culture, and venture capital. The gap is operating a small field-service firm and closing small-company deals. The plan is a co-founder who has run or sold such a firm, and an industry president per vertical.