Self-improving invoice extraction pipeline
Finance operations, Procurement
An invoice extraction pipeline on a fine-tuned Gemini model that retrains every week on newly processed and corrected invoices, reaching 96% extraction accuracy.
AI Automation & Agents· Web Application
Estimate AI
An extraction platform that trains a dedicated model for each vendor, turning inconsistent construction estimate PDFs into structured, ERP-ready data.
- Field accuracy on the evaluation set
- 96%
- A fine-tuned model for each supplier's estimate format
- Per vendor
- Estimators train and deploy new models without an engineer
- Self-serve
- Structured output for estimating and ERP systems
- ERP-ready
The starting point
The challenge
Construction teams receive quotes, scopes and estimates as PDFs, and every vendor lays them out differently. Keying them in by hand was slow and error-prone, and generic OCR broke on the variety of tables, units and totals, so the data never reached estimating or ERP systems cleanly.
The solution
What we built
Built a Django and React platform where estimators create a model for each vendor, then upload example PDFs alongside the structured output they expect.
Automated dataset preparation: examples are parsed, cleaned and paired into training data, and checked before any fine-tuning job starts.
Ran fine-tuning as background jobs on Celery and Redis, so new vendor models train without blocking anyone using the app.
Trained models extract structured data from new estimates, mapped to the fields that estimating and ERP systems expect.
Architecture
How the system works
- 01Vendor examplesEstimate PDFs with expected output
- 02Prepare datasetParse, clean and pair examples
- 03Fine-tuneBackground job per vendor model
- 04EvaluateScored against labelled estimates
- 05ExtractNew PDFs become structured data
- 06ERP & estimatingClean, reusable output
Engineering
How it was engineered
Built evaluation sets of labelled vendor estimates, so every vendor model is judged on documents like the ones it will see.
Scored every model iteration against those sets before it could replace the version in production.
Kept evaluating on live documents after launch, to catch drift as vendors changed their layouts.
Scope
What was delivered
01
Vendor model management interface
02
PDF parsing and dataset preparation pipeline
03
Background fine-tuning job scheduler
04
Structured extraction service
05
Per-vendor evaluation sets
Gallery
Screens and diagrams
Technology
Stack
DjangoReactOpenAI fine-tuningCeleryRedisPDF parsing
More work
Related case studies
Field-to-finance ETL for a construction company
Construction, Finance operations
An Apache Airflow pipeline that moves the hours, machine usage and materials logged in SmartDok into Duett for payroll, timesheets and invoicing, with no manual re-entry.
ERP documentation from interface screenshots
ERP, Enterprise software
An AI workflow that reads ERP interface screenshots, recognises the screens, fields and workflows on them, and drafts user documentation.
Need something similar?
We can scope and build a system with comparable capabilities, shaped around your own workflows, data and integrations.