Document intelligence for an Australian state government agency, turning survey PDFs into audit-ready Salesforce records

The challenge

The agency’s compliance register depends on inspection and survey reports produced by several external consultancies. The reports themselves were sound. Everything after them was the bottleneck.

Each consultancy used its own table layouts, column headers, field ordering and labelling conventions. Staff re-typed the contents into a shared spreadsheet page by page. A single survey took hours. No one trusted the spreadsheet without opening the source PDF again, and there was no fast way to prove where a given value had come from.

The register data belongs in Salesforce, where it drives the agency’s compliance workflow. Manual transcription sat between the two, absorbing skilled time and introducing exactly the kind of error a compliance record cannot carry.

What we built

We built a document intelligence platform, not a chatbot. It turns consultant survey PDFs into structured, audit-ready register data aligned to the agency’s Salesforce schema.

• Ingestion. PDF parsing in accurate table mode with cell matching. Two parsers ran side by side for two months behind a consensus engine before we retired the weaker one, so the choice was made on evidence rather than preference.
• Extraction. A twelve-node extraction graph and a three-node source-processing graph, orchestrated in LangGraph. Extraction runs per table row rather than per document. Bulk document prompting read fluently and failed silently on merged cells and page breaks, which is the failure mode a compliance record can least afford.
• Format handling. Schema inference with cached per-consultant format profiles. Five consultant report layouts run through one configurable parser. Adding a consultancy is a configuration change, not a new parser.
• Provenance. Every extracted cell is traceable to its source page and bounding box in the original PDF.
• Human review. A web review surface presents the register as a wide editable grid, with each value paired to its source location. Nothing reaches Salesforce until a reviewer has confirmed it.
• Domain rules and validation. The agency’s own data entry rulebook is encoded as defaulting and formatting rules. Output is validated against Salesforce picklists, picklist dependency triples and record validation rules before export.
• Write-back to Salesforce. Bulk-upload-compatible CSV and Excel exports, plus live upsert into the agency’s Salesforce org through a Connected App and OAuth, keyed on an external ID. Dry-run preview, per-record error surfacing and sync tracking are built in.
• Operations. Passwordless sign-in, queued bulk upload with per-user job queues, streaming job status, and a dedicated UAT environment.

Data sovereignty was a design constraint from the start. Inference runs on Australia-resident model profiles in an Australian workload region, and observability was moved onshore during the build. Spend guardrails were designed before the first production extraction: IAM-enforced hard caps, per-model rate limiters, prompt caching and a kill switch. Spend has stayed inside those caps.

How we worked

The MVP was built between February and May 2026. Accuracy work ran as measured campaigns against manually validated ground truth drawn from real consultant surveys in several formats, not as ad hoc tuning.

Ground-truthing against the agency’s live Salesforce sandbox rather than against a legacy export was a turning point. It corrected the target schema early and reshaped the field alignment work before it hardened.

Agency domain reviewers then tested the MVP against ten real consultant surveys. Their findings were reconciled into a phased forward plan, proposed in June 2026. Phase 1 began in July 2026, covering extraction architecture, the Salesforce write-back integration, QA, UI and UX review, and release and environment ownership. Reviewers also handed over the agency’s internal data entry conventions, which became a workstream of its own.

Delivery runs on Bitbucket Pipelines CI/CD into AWS. As of September 2026 the platform is in user acceptance testing on a dedicated environment, with the production environment being stood up.

Outcomes

• Replaces manual page-by-page transcription of consultant survey reports with a reviewed, structured extraction pipeline, now in user acceptance testing.
• One configurable pipeline handles five different consultant report layouts. New consultancies are onboarded by configuration.
• Every extracted value carries cell-level provenance back to its page and position in the source PDF. The bounding box, not the model’s accuracy, was what the compliance team trusted first.
• Live Salesforce upsert built and delivered into test in September 2026, closing the path from PDF to compliance record without a manual re-keying step.
• An optimisation campaign on a 190-page benchmark survey cut end-to-end processing from about three hours to roughly half an hour, and cut per-document inference cost by close to ninety per cent.
• Runs within enforced monthly cloud spend ceilings, with a quantified and capped worst-case runaway scenario.
• All customer data and inference stay onshore in Australia.
• Accessibility reviewed to WCAG 2.1 AA.
• The figures above come from controlled build benchmarks. No production usage metrics exist yet and none are claimed here.