Precision components manufacturerSample
Automating quality reports and certificates for a precision manufacturer
A precision components manufacturer supplying aerospace and medical customers put together every certificate of conformance and inspection report by hand, from CMM outputs, paper inspection sheets and supplier certificates. We built a system that pulls the data out automatically, so QA engineers check values instead of typing them, and produces the customer paperwork in minutes where it used to take hours.
- Client
- Precision components manufacturer
- Industry
- Manufacturing & Industrial
- Year
- 2025
- Duration
- 5 months
- Team
- 4 people
Illustrative figures for a representative engagement.
The challenge
Where things stood when we arrived.
Every shipped batch needed a certificate of conformance and, for many customers, a full inspection report with measured values against drawing tolerances. The data lived in three places. Coordinate measuring machines produced CSV exports, the shop floor produced inspection sheets (handwritten or printed), and suppliers sent PDF material and treatment certificates. A QA engineer collated all of it into Excel and Word templates for each batch.
It took two to three hours per batch, and finished parts routinely sat waiting a day or more for paperwork. Transcription errors (a wrong decimal, a missed feature) were caught by customers often enough to be embarrassing, and occasionally to trigger a formal non-conformance.
The company ran an ERP that held orders and routings but nothing about measurements, and its network policy meant machine data could not leave the shop-floor segment directly. Anything we built also had to meet AS9100 traceability requirements, which means every value on a certificate has to be attributable to a source record.
The solution
What we built, and how it fits together.
We built a document and data extraction pipeline in Python. CMM exports are parsed directly and matched to drawing features. Inspection sheets and supplier certificates are scanned or uploaded, run through OCR and then through an LLM extraction step that maps handwritten values, material grades, heat numbers and signatures into a structured schema, with a confidence score on every field. Anything below threshold, or failing a validation rule such as a value outside tolerance or a heat number that isn't on the purchase order, goes to review.
A web tool presents each batch to a QA engineer as one review screen, with the extracted values next to the source image and any discrepancies highlighted. Approving or correcting a value is one click. Approval generates the certificate of conformance and the inspection report from per-customer templates, and stores the documents, the extracted data and links to every source file against the batch. Corrections feed back into the extraction rules and the prompt evaluation set.
The system runs as containers on an on-premise server that bridges the shop-floor and office networks, with a small AWS footprint for the LLM API calls through an allow-listed egress. No drawings or customer identifiers leave the site. Order and routing data comes from the ERP by API, and the ERP is updated with certificate references when a batch is released. Deployments are scripted and monitored, and can be rolled back without stopping production.
Approach
The order we did things in.
Technology
- Python
- OpenAI / Anthropic APIs
- TypeScript
- React
- Next.js
- Node.js
- PostgreSQL
- Docker
- AWS
- Grafana / OpenTelemetry
- 01
Sampling the paperwork
Two weeks with QA collecting six months of certificates, inspection sheets and CMM files, timing how long each batch took, and building an evaluation set of a few hundred documents with known correct values.
- 02
Tuning extraction against the evaluation set
No screens were built until the pipeline was right. We worked through the evaluation set until field-level accuracy and confidence calibration were good enough that reviewing would take a fraction of the time transcribing did.
- 03
Review tool and parallel running
QA engineers used the review tool alongside the manual process for six weeks, comparing generated certificates against the hand-made ones and refining templates and validation rules as they went.
- 04
Moving customers across, then handing over
Generated paperwork replaced the manual process for each customer in turn. The quality manager got an audit walkthrough and runbooks, and the IT team was trained on deployment and monitoring.
Results
What changed.
- QA engineers save roughly 25 hours a week across the team, most of it time that went on transcribing values
- Transcription errors found in customer paperwork fell by around 80%, and the rest are caught in review before release
- Certificates and inspection reports go out the same day a batch passes inspection, where before they took two to three days
- Every value on every certificate links to its source CMM file, inspection sheet image or supplier document, which is better traceability than the manual process ever had
- The quality team maintains templates and validation rules itself, and the evaluation set catches regressions when extraction prompts or models change
“Our engineers used to type numbers off photocopies. Now they check measurements, which is the job we hired them for.”
Sample quote. Replace with an attributed client statement.
Similar situation?
Tell us about it. If we're not the right people for the job, we'll say so.