Orbit Supply · 2026
Extracting tender line items for Orbit Supply with citations a buyer can check
Orbit's buyers rekeyed 40,000 pages of tender documents a month, and the errors only showed up at award. We built schema-constrained extraction with page-level citations and deterministic validators, so buyers verify highlighted figures. Intake dropped from 26 minutes to four per tender.
- Public Sector
- 6 months
- 5 engineers, 1 ML engineer, 1 designer
- 4 minutes
- Intake per tender, down from 26 minutes
- 98.2%
- Field accuracy on a 1,200-document gold set
- GBP 0.04
- Model cost per document processed
The challenge
What was wrong
Orbit's buyers rekeyed line items from about 40,000 pages of tender documents a month, arriving as native PDFs, scanned Word files and nine incompatible framework templates. Rekeying took 26 minutes a tender and introduced quantity and unit errors that only surfaced at award, when they were expensive. Public procurement rules mean a named officer stands behind every figure, so anything producing numbers without a traceable source was unusable.
The approach
What we did
Extraction returns a strict schema with a page and bounding box citation for every field, so a buyer verifies a highlighted number instead of reading the document again. Deterministic validators then run over the output: unit conversions, line totals against stated sums, date sanity. Fields that fail validation or fall below confidence route to review, which is 31 per cent of documents. We used a mid-sized model with retrieval rather than the largest available, holding cost to GBP 0.04 a document at 12 seconds.
“Every figure links back to the page it came from, with the box drawn on it. That is what let our legal team sign off on using a model in procurement at all.”
Stack
What it runs on
- Python
- FastAPI
- Anthropic Claude API
- Azure Document Intelligence
- pgvector
- PostgreSQL
- Next.js
- OpenTelemetry
Services
Practices involved
AI & Data
LLM Integration
Wiring language models into systems that already have owners, SLAs and auditors.
AI & Data
AI Development
Applied AI systems built to run in production, not to demo well.
Engineering
API Development
Contract-first interfaces with versioning, quotas and deprecation rules agreed up front.
Engineering
Enterprise Integration
Contracts, idempotency and replay — the unglamorous parts that keep 40 systems in step.
More work
Other engagements
Meridian Exchange · 2024
Bidding engine rebuild
Meridian was losing bids to lock contention in the closing seconds of every auction, and disputes over bid order had to be settled manually. We rebuilt the engine around per-lot single-writer partitions and an append-only log, trading cross-lot transactions for provable ordering at 1,800 bids per second.
Kestrel Industrial · 2023
Shop-floor system modernisation
Four plants, 300 machines and a VB6 application nobody could test safely. We replaced it one production cell at a time behind a routing gateway, keeping Oracle as the book of record until the final cell was live. Total unplanned stoppage across the entire programme came to 11 minutes.
Northwind Commodities · 2025
Incremental end-of-day valuation
A seven-hour overnight batch left Northwind's traders working from yesterday's P&L whenever a curve was corrected. We replaced it with an incremental graph over versioned inputs, cutting the full run to 34 minutes and a curve correction to 90 seconds, with 12 months of valuations replayable to the input.