AI Enterprise SolutionsGPT & Claude

TRACE

TRACE is designed for construction estimators, quantity surveyors, and procurement teams who must reconcile technical specifications, Bills of Quantities (BOQs), revisions, and supplier quotations. Today, these checks often require manually comparing documents with different formats and item descriptions, making it easy to miss scope gaps, quantity or unit mismatches, and specification conflicts. TRACE extracts document content, identifies potential inconsistencies, presents evidence from the source documents, and helps the estimator prioritize issues for human review. It supports the review process; it does not approve tenders or replace professional judgment.

Open the demo Video GitHub

Preliminary score

66.5/ 100Promisingaverage of 2 judges

Value for the user19.0 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI22.5 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing10.0 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility9.0 / 15
Data requirements, running costs and a clear next step.
Originality6.0 / 10
A meaningful difference from common solutions.

What the judges wrote

Claude Opus 5.5Preliminary65/100
Strengths
Your problem framing is concrete: estimators and QA/QC staff reconciling specs, BOQs, revisions and supplier quotes. Your findings are specific, such as C30/37 versus B30 concrete and an unapplied 850 m² yard sub-base revision. The architecture is well thought out: browser-side extraction, a strict JSON-schema LLM call, Zod re-validation with rejected findings surfaced, a banned-claims filter, and deterministic TypeScript for all differences and cost impacts. Your write-up is honest about limitations, including synthetic data only, no recall or accuracy figures, no scanned PDFs and the 60k-character cap.
To improve
Your testing is one synthetic package with no ground truth, so build a labelled set with deliberately planted discrepancies. Then report catch rate, false positives and missed issues, with at least one concrete failure example. Time a manual review of the same package against TRACE to give the efficiency metric your deck proposes. Add a per-audit running cost (tokens and price) and a named pilot partner or real tender dataset as your next step. Also align the target user between the deck (QA/QC and certification engineers) and the submission (estimators and procurement).
GPT-6 AstraPreliminary68/100
Strengths
Your tender-review workflow addresses a concrete task for estimators and procurement teams, connecting discrepancies to source excerpts and reviewer actions. Your architecture sensibly separates AI interpretation from deterministic arithmetic and keeps final decisions with the reviewer. Your synthetic test reports specific discrepancies and clearly distinguishes preliminary cost exposure from proven savings.
To improve
Your next evaluation should use independently labelled tender packages and measure false positives, missed issues, citation validity, and review time against manual review. Make the synthetic run reproducible with expected outcomes for all 16 checks and the quantities and rates behind the ₼39,625 estimate. Demonstrate that quotations and page/row references are checked against source text, beyond schema validation and prompting. Add measured cost and latency per audit, plus a retention and access-control plan for confidential client documents.

The team

Laman Asgarova
developer
Pari MammadovaLead
developer
Natig Alizada
developer

Disclosure

Required by the rules. Judges verify it against the repository.

Models: GPT-6, AI model accessed through the Lovable AI Gateway
Data: Synthetic construction tender documents and sample BOQ data for demonstration; no real partner dataset is claimed.
Components: TanStack Start, React, TypeScript, Tailwind CSS, shadcn/ui, Zod, browser-side PDF/spreadsheet/document extraction libraries, and Lovable AI Gateway. Exact dependencies to be confirmed in the repository.
Built during the event: Confirmed

Quality testing

Testing approach: TRACE was tested using a synthetic construction tender package containing a technical specification, BOQ, supplier quotation, and revision addendum. The audit performed 16 checks and returned 8 potential inconsistencies: 7 high-severity and 1 medium-severity finding, alongside 8 passed checks. Observed results: The system identified specification mismatches involving concrete strength, roof deck thickness, roof panel requirements, and feeder cable size. It also flagged revision-related discrepancies involving an additional 850 m² of yard sub-base and two loading doors, as well as potential gaps in sprinkler-system scope and door safety accessories. Findings included source-document references and recommended actions for human review. Commercial impact: TRACE estimated ₼39,625 in potential additional costs for the missing yard sub-base quantity and two loading doors. This is a preliminary estimate based on stated unit rates and requires verification; it is not a complete estimate of all potential cost differences. Limitations and comparison: The test demonstrates evidence-linked discrepancy detection on a synthetic example, but does not establish overall accuracy, recall, or time savings. The eight passed checks were not individually detailed in the report, and the findings have not all been independently verified. Compared with manual document-by-document review, TRACE aims to consolidate potential inconsistencies and their supporting evidence into one review workflow. Further testing against manually labelled tender packages is needed to measure false positives, missed issues, and performance on real-world documents. All findings require human verification.

Worth 20 of the 100 points.

Running it

Open TRACE: Launch the application using the provided project link. Upload documents: Add the relevant BOQ, technical specifications, and supporting files. For testing u can add provided files from the github link src/test-data folder. Run the AI audit: Start the analysis to identify discrepancies, missing information, and potential compliance issues. Review findings: Inspect each flagged issue and its supporting evidence. Make a decision: Verify the findings and determine the next action.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.