Munta AI for industrial decarbonisation · United Kingdom

Trustworthy AI for high-stakes engineering

An AI engine for industrial decarbonisation — that proves its own numbers before they reach your board.

Munta models a site's heat and cooling hour by hour, ranks the costed decarbonisation pathways — heat pumps, heat recovery, electrode boilers, thermal storage — and holds every figure to an audit standard before it ships. The result is a review-ready study a capital committee can act on, in about three weeks.

Reconstructed 3 completed IETF heat-pump projects from public data alone under a no-retuning rule
8,760
hours modelled per site — a dispatch model, not a spreadsheet average.
100%
of figures traceable to their derivation; unverified numbers never ship.
~3 wks
from complete inputs to a board- and ESOS-ready draft.

The engine

An AI system built to be trusted with engineering decisions

Most AI gives you an answer. In engineering, an unchecked answer is a liability. Munta pairs frontier language models with a physics-based simulation engine — and a verification discipline that treats the AI's own output as a draft to be audited, not a result to be believed.

Models

Physics + AI, not prose

An hour-by-hour thermodynamic model of the site — steam, hot water, process cooling — drives the analysis. The AI reasons over real engineering, from heat-pump cycle design to whole-site energy dispatch.

Ranks

Costed, decision-ready pathways

Every viable route is costed and ranked against your payback gate — capital cost, carbon saving, NPV, IRR and simple payback — with Monte-Carlo uncertainty on the headline case.

Verifies

It surfaces its own errors instead of hiding them

Before a figure reaches you, it is held to an audit standard: every number traces to its derivation, and unverified or unsupported values are stopped, not shipped. On one benchmark the engine under-recommended a real project — and its own checks caught it, before a human did. That is the difference between an AI you can put in front of a board and one you cannot.

The deliverable

The Stage-1 Scoped Pathway Study

One site, one fixed fee. The deliverable a capital committee can act on — and hand to its ESOS Phase 4 assessor.

An 8,760-hour model of your site

Steam, hot-water and process-cooling loads built from your own 12 months of energy and process data — not a spreadsheet average, an hour-by-hour dispatch model.

Ranked, costed pathways

Every viable route — waste-heat heat pumps, heat recovery, electrode boilers, thermal storage — with capital cost, carbon saving, NPV, IRR and simple payback, ranked against your gate.

Cash vs appraisal-carbon, split out

Monte-Carlo uncertainty on the headline pathway, and an honest split showing where a case rests on fuel-bill cash versus a carbon-appraisal value — so no one is surprised at the capital gate.

Board- and ESOS-ready

A review-ready draft with full calculation provenance — every figure traceable to its derivation — suitable as direct input to your ESOS Phase 4 action plan (due 5 December 2027).

Fixed scope. Fixed fee. And a warranty: if a suitably qualified engineer on your side finds a material error in a calculation derivation, we correct it at no charge — and if it changes the study's conclusion, you don't pay for that site.

Pricing scoped per engagement · discussed on the call

Proof, not promises

Measured against real, completed projects

Before it models your site, the engine was tested against three completed, publicly documented industrial heat-pump projects — each rebuilt from the public figures alone, then run unmodified.

Dale Farm Cookstown dairy · £2.85m built

Rebuilt the completed heat-pump decision from GOV.UK figures. Matched the technology and the capital-cost scale, and landed a carbon saving inside the published band. On the first run it under-recommended the project — the engine's own validation caught that against the real outcome, fixed it, and re-ran to confirm.

Capital scale£2.2m vs £2.85m
Technologymatched
Found → fixed → verified
Heinz Kitt Green · heat pumps

Reconstructed the site's carbon arithmetic from public data and reproduced the published saving to within about 16%. The selected pathway stood up on fuel-bill cash and grant alone — before any carbon-appraisal value was counted.

Carbon vs publishedwithin ~16%
Cash + grantpositive
Inside published band
ABP Shrewsbury · beef plant

The like-for-like reconstruction reached 62.6% gas displacement against a published >80% aspiration — inside the acceptance band, and honestly short. Rather than tune the inputs to match, the engine diagnosed exactly where and why the gap sat.

Gas displaced62.6% vs >80%
Shortfalldiagnosed
Honest miss, diagnosed

The honesty is the point. A study you can put in front of a board has to show where it is wrong, not only where it is right — the same assurance you would demand of a junior engineer's draft before it goes near a client.

The process

How it works

Three steps, about three weeks from complete inputs.

Step 01

You share the data

Twelve months of gas and electricity use, a load description, and boiler and chiller schedules. One nominated technical contact.

Step 02

The engine models, ranks & verifies

An hour-by-hour dispatch model, a ranked set of costed pathways with Monte-Carlo uncertainty, and the cash-versus-carbon split — each figure audited before it lands in the draft.

Step 03

You get a decision

A review-ready study — board-ready and ESOS Phase 4 input — with every figure traceable to its derivation, plus a walk-through call.

Who's behind it

Engineering judgement, built into the software

Munta is built by a mechanical engineer working in industrial energy and high-voltage utility software — the engine, the methodology and the study are one person's work, deliberately kept rigorous.

The background is thermofluids and simulation — 100+ industrial CFD and energy studies for operators across power, oil & gas and offshore wind — now applied to the harder problem: making an AI system trustworthy enough to sign off. Its defining discipline is provenance: every number in a study traces to its derivation, and the engine's own checks are built to surface its errors rather than hide them — which is exactly what the benchmarks above demonstrate.

Request the full methodology →

Have a published IETF project, or a site facing ESOS Phase 4?

Tell me the site and I'll show you what the engine reconstructs for it — a 20-minute call, no obligation. If a benchmark of your own published project would be more useful, I'll run that too.