TRL 3 — experimental proof of concept Experimental prototype developed by Syrus Industry Candidate solution for COSMIC Challenge 29

Energy bills are hard to read. We are building software that reads them.

ClearEnergy AI is an experimental prototype that reads energy bills, extracts the values that matter, and explains them in plain language — with every number traceable back to the document it came from.

See what exists today See what is planned

Where this project actually stands

This page describes an experimental prototype at TRL 3. It is not a product, it is not commercially available, and it has not been validated in an operational environment.

Syrus Industry is a candidate for COSMIC Open Call 2, Challenge 29. No funding has been awarded, and no affiliation with COSMIC, Horizon Europe or the European Union is claimed or implied.

The sections below separate what the prototype does today from what is proposed for future work. Nothing in the second list exists yet.

The current prototype

A working web prototype, developed by Syrus Industry with its own resources, operating on a controlled dataset of real Italian energy bills belonging to the company itself.

30
real Italian PDF bills
2
suppliers tested
2
energy types (power, gas)
2023–2026
period covered

What the prototype does today Verified on the dataset described above.

  • Processes 30 real Italian PDF bills from two suppliers: ACEA Energia and Eni Plenitude.
  • Covers both electricity and gas supplies, spanning November 2023 to April 2026.
  • Extracts the main values from each bill: billing period, consumption, unit cost and total.
  • Displays the results as figures, tables and charts in a web interface.
  • Compares periods, costs and consumption against each other, month by month.

What the prototype does not do Stated explicitly, to avoid any misreading of the list on the left.

  • It is not adapted to Portuguese bills. It has only been exercised on Italian ones.
  • It is not integrated with CTF6, and no such integration has been attempted.
  • Photo support is not validated. The dataset consists of native PDF files only.
  • The 90% extraction accuracy target is not demonstrated. It is an objective of the proposed work, not a measured result.
  • The dataset is a business supply, not a sample of domestic households.

How it is built

One design decision shapes everything else: the language model is never allowed to produce a number.

Every economic and regulatory figure is computed by deterministic code from structured data. The language model receives results that have already been calculated, and may only describe them.
StageWhat happensWho does it
IngestionFormat and MIME check, SHA-256 hash, deduplicationDeterministic code
ClassificationSupplier and document layout identified from contentDeterministic code
ExtractionValues located by position and anchor labels in the documentDeterministic code
ValidationSchema check, internal consistency, per-field confidence scoreDeterministic code
CalculationCosts, comparisons, indicators — from versioned rules and sourcesDeterministic code
ExplanationDescribing results already computed, in plain languageLanguage model

Proposed work — planned and to be validated

None of the following exists today. This is what Syrus Industry proposes to build and validate under COSMIC Challenge 29, starting from the prototype described above.

  • Processing of Portuguese energy bills.
  • Coverage of at least three major Portuguese suppliers.
  • OCR of photographed bills.
  • Extraction accuracy of at least 90%, measured on a controlled dataset.
  • A Portuguese data model.
  • A tariff and regulatory engine with dated, verifiable sources.
  • Deterministic checking of individual bill line items.
  • Anomaly detection on consumption and costs.
  • LLM explanations with sources and confidence scores.
  • CO₂ estimation from published emission factors.
  • Personalised recommendations.
  • Accessibility for users aged 55 and above.
  • API integration with CTF6.
  • A pilot with approximately 40 Portuguese households.
  • Progression from TRL 3 to TRL 6–7.