Concept · Architecture · Build
Engineering Estimation Engine
Reads a FEEED scope package and produces a calibrated estimate package — basis-of-estimate, confidence band, full audit trail back to the source paragraph. Replaces the spreadsheet-and-judgement loop most operators still run.
- Stack
- HTML SPA, IndexedDB, Local LLM, Retrieval-augmented norms
- Outcome
- Two-day estimating cycle compressed to ninety seconds
- Year
- 2026
The bottleneck
Engineering estimation in EPC and PMC is still a workflow where two senior engineers sit with a scope statement and a library of historical norms, argue for two days, and produce a number with a credibility band of plus or minus thirty percent. The norms libraries are PDFs. The scope is a Word document. The judgement is in someone’s head.
What the engine does
- Phase 0 — extraction. A small open-weight model running locally on a laptop reads a FEEED scope document and emits structured tags against a schema — counts, totals, density measures, narrative-complexity flags. Nothing leaves the laptop.
- Phase 1 — calibration. Tags map onto a hybrid norms engine grounded in industry-standard cost-class methodology and stage-gate process discipline, plus operator-specific maturity calibration. Output is a structured estimate package with a stated confidence band.
- Phase 2 — output. A basis-of-estimate worksheet, a CTR sheet ready for tender, and a JSON record of every assumption made. Every value carries a citation back to the source paragraph.
Why this matters
The estimation side of EPC project work is the last domain still defended by “it takes years of experience.” It does — and that experience can now be encoded, repeated, and audited. The engine does not replace the senior engineer. It removes the two days of arguing and gives them a structured starting point in ninety seconds.
What was shipped
- Phase 0 scope-extraction catalog, sealed against drift, running on three case sets
- Phase 1 hybrid calibration engine, stress-tested across a wide range of program scales
- All engineering modules wired to the same maturity matrix from concept through detailed design
The next phase is bid-volume calibration — feeding actual program outturns back to refine the norms automatically.