On Hong Kong public electrical and mechanical works, a contractor cannot install a luminaire, cable or distribution board until the material is proven compliant. Each batch arrives as a pack of datasheets, test reports and certificates that a government reviewer must check against the General Specification, the project's Particular Specification and the works checklist.
One submission means 22 checkpoints found by hand across six documents and 200+ pages, including scans and dense tables. It is slow, a clause is easy to miss, two reviewers grade the same risk differently — and where an answer was found is never written down, so nothing can be re-checked later.
GPTBots turns both specifications into one executable rule set, merged with the reason each rule governs. Every submission is parsed — OCR for scans, structure preserved in tables — then audited rule by rule. Each verdict carries the wording it was based on and the page it came from, and the reviewer confirms or overrides it.
Reviewers confirm findings instead of hunting for them, so the human time per checkpoint falls sharply. Every conclusion is defensible to a page, risk grading stops drifting between people, and hundred-page packs stop falling back to manual review. Live in production with a Hong Kong government department today.
Every batch of contractor material arrives as a stack of datasheets, test reports and certificates that has to be checked, clause by clause, against two specifications and a checklist.
One material pulls in the general spec, the project spec, the checklist, product literature, test reports and certificates — across PDF, Word, Excel and scans.
Submissions run to hundreds of pages of dense tables, scanned pages and test data. Handed to a model whole, they lose table structure, drop parameters or time out.
The project spec may add to, override or contradict the general spec. What a reviewer needs is the final rule — and why that rule won.
“Compliant” is not enough. The reviewer has to see which rule, which file, which page — and jump to the highlighted wording.
Risk grading, acceptance method and the wording of an opinion drift between people, which shows up later in reporting and internal re-checks.
These verdicts gate works acceptance and carry compliance liability. A missed clause is not a bad search result — it is a material installed on a public project that should not have been.
Not a document chatbot. Built for the material submission and approval process on electrical and mechanical works, the system carries a versioned, executable rule set, audits each submission against it, and produces a reviewable record where every finding is traceable to source.
Stops opening four documents to check one parameter. Starts from a drafted conclusion with its evidence attached, and spends the time on judgement instead of retrieval.
Gets one house standard for risk grading, acceptance method and report structure, plus a dashboard showing what is waiting on a person and what is waiting on the system.
Finds out precisely which requirement is unevidenced and which document would close it — instead of a general rejection and another submission cycle.
Four stages exist so the system holds the rules. Three exist so a human can defend the verdict afterwards.
Specifications stop being documents to read and become a versioned rule set the system can execute. The general and project specs are merged into one final set, with the reason each rule governs.
Every rule in the applied set is checked against the parsed submission. Each produces a verdict, a written opinion and a risk grade — and where nothing evidences a rule, the system says so rather than asserting it.
Each conclusion is mapped back to file, page and passage, and rolled into a report. The reviewer confirms, edits an opinion in place, or overrides a finding — and the record keeps what a person changed.
These carry over to any “specification + evidence + accountable verdict” scenario, not just electrical and mechanical materials.
Format problems are solved before the model sees anything, so parsing accuracy and reasoning accuracy can be diagnosed separately.
Fields, types, risk levels and error returns are agreed first; the model produces to that contract so the UI and report can consume it.
A conclusion arrives with its source, so a human can verify it in seconds. This is the principle that decides whether the customer trusts the system.
An extra engineering layer — split, parallelise, merge, correct page numbers — is always preferable to silently dropping content.
Pick a material, choose which documents the contractor attached, and run it. The verdicts are computed from what you attached — leave a certificate out and watch which checkpoint fails.
Captured from the environment your customer can open today. Everything below is real output, not a mock-up.
Projects, materials, pending items and active rule count, with running and queued audit tasks alongside project and material status distribution — so a supervisor can see what is waiting on a person and what is waiting on the machine.
Each row carries a requirement, the acceptance method that would let it pass, and a risk level. Rule sets are versioned with change notes, and the source file stays downloadable — so an auditor can always ask where a rule came from.
22 checkpoints, filtered by compliant, partially compliant, non-compliant or pending evidence. Each shows the wording the system extracted, the page chips it came from, and an editable audit comment. Click a chip and the viewer jumps to that page of the 107-page submission with the passage highlighted.
Above a configured page count a submission is split into chunks, audited separately and merged back. Audit and rule-generation concurrency are tunable to the deployment's capacity, and ad-hoc reviewer requirements can be synced back into the project rule set as a new version.
The manual effort moves from finding and collating evidence to checking a conclusion that already has its evidence attached.
Standards, parameters, certifications and acceptance methods are covered systematically rather than from a reviewer's recall.
Fields, risk levels, acceptance methods and report structure stay consistent between people, which makes internal re-checks tractable.
Every verdict resolves to a rule, a file and a page — the evidence standard government and engineering compliance work actually requires.
Hundred-page packs run to completion instead of falling back to manual review because parsing failed or content was truncated.
Front end, back end and AI capability are built and maintained by one team, so change requests and defects have a single owner.
The saving is not a percentage off the old process — it is the same checkpoints, counted twice. Once at the minutes it takes to find and judge each one from scratch, once at the minutes it takes to verify a verdict that already carries its evidence. Defaults are our estimate, not measured results from this deployment; change any of them in the meeting and the answer moves with you.
Reviewer hours redirected from searching to confirming. The most defensible pool, because the customer already knows all three inputs.
The contractor is told which requirement is unevidenced and which document closes it. Each cycle removed is calendar time off the works programme, not just review time.
A clause missed at submission surfaces at acceptance, with the material installed. Usually the largest pool and the hardest to quantify: ask what one such event has cost them.
What goes in, what comes out, who can sign in and where it runs — the questions an IT reviewer asks before a business owner is allowed to say yes.
ARM (GX10) and x86 (RTX 6000 Pro); the GPU environment is faster and parses more accurately.This deployment is document-in, record-out — it does not currently sit behind a document management system or ERP connector.
Where a customer needs submissions pulled from an existing system, or verdicts pushed into one, that is scoped as integration work on top of the GPTBots platform rather than something already shipped. Say so plainly in the room — it is a shorter conversation than walking it back later.
If a competitor is pitching “chat with your documents”, these are the four questions that separate the two.
A summary cannot be signed. Every verdict here carries a rule number, a file, a page and the original wording, and the reviewer can click straight to the highlight. That is the difference between something interesting and something a government department can act on.
The first version of this system only reported what the project spec changed. Reviewers needed the final executable set and the reason each rule governs. That rework is already paid for — the merging workflow ships with change type, confidence and source reference.
OCR for scans, code components for tables, and a split-parallelise-merge path with page-offset correction for oversized PDFs. Most failures in this category are parsing failures wearing a reasoning costume.
Front end, back end and AI capability are built and maintained by us. There is no seam between the agent vendor and the application vendor for a defect to fall into, and the customer has one number to call.
Nothing in the seven stages is specific to lighting. Moving to another field of work means loading its specifications and its checklist — the parsing, the rule merging, the audit and the evidence trail are already built and already in production. A second project starts roughly where this one finished.
Every one of these is the same problem wearing different clothes: a rulebook on one side, a pile of submitted evidence on the other, and a verdict someone has to stand behind. If a scenario fills in these four columns, the engine already fits it.
Only the first row is built. The rest are scoped as rule-library work on the existing engine — bring one to the project team before promising a timeline.
Three questions that qualify a scenario inside a first meeting. All three true is a strong match.
Written standards exist in layers or versions — national against company, general against project-specific.
One review means reading several long documents, and most of the time goes into searching rather than deciding.
The conclusion has to state its basis, because someone may re-check or audit it later.
Yes — in production with a Hong Kong government department, with all seven stages in real use.
Typically tens to hundreds of pages. Beyond a configured page threshold the file is split, parsed in parallel and merged back together.
Yes. Scans go through OCR; tables are restructured by code components before the model sees them.
Three constraints: rules are produced to a structured contract rather than as free text; every conclusion carries a rule number, file, page and original passage you can click back to; and the system is positioned as first-pass review — the reviewer signs the conclusion.
A separate merging workflow labels each rule added, overridden, conflicting or unchanged, and returns the reason, a confidence value and the source reference.
This project was deployed inside the customer's own environment, verified on both ARM (GX10) and x86 (RTX 6000 Pro) hardware.
Both environments run the full flow. The GPU environment is noticeably faster and parses more accurately, so plan the target environment early and validate against it.
The capability is packaged as a WorkSpace extension application, so the work is rule and scenario adaptation rather than rebuilding workflows. Confirm timeline with the project team before quoting one.
A two-minute recorded run through the live system, on the real government deployment.