Custom Tools (Industrial — Capacity Unlock)
Same team, same trucks, a different tool in their pocket. Technicians now close five service jobs per day instead of three. A 67% capacity lift without a single new hire, driven by Claude Vision photo-to-part recognition.
Servikom Kompresorji services industrial air compressors across Slovenia. The constraint on the business was never mechanical skill — it was parts identification in the field. Every repair started with ten to fifteen minutes of flipping through paper catalogs and calling back to the depot. We replaced the catalog with a mobile tool that recognizes any part from a photograph, and the daily throughput of the service fleet shifted from three jobs to five.
Servikom was turning down work. Service slots were full, quotes were going to competitors, and the obvious fix — hire more technicians — was blocked by training lead time (roughly six months to certify a new field tech on compressor service) and by the wage cost of a second shift.
The alternative was to lift the output of the existing team. The team's time map was clear: billable work was the mechanical repair itself. The dead weight was parts lookup, which was eating 10-15 minutes at the front of every job because the catalog lived in a mix of paper binders, supplier PDFs, and scattered emails. The tools were the bottleneck, not the people.
A mobile-first parts tool the field techs carry on their phones. Three ways in: text search across a normalized catalog, QR scan on equipment nameplates, and — the feature that actually moves the needle — photo-to-part recognition powered by Claude Vision.
A tech snaps a photo of a compressor, a component, or a worn-out part. A Python pipeline normalizes the image, runs OCR on any visible serial, and hands the composite to Claude Vision with the structured catalog as grounding context. The tool returns a ranked list of candidate parts with confidence scores, cross-referenced to current stock and supplier lead times.
The catalog sits in SQLite with FTS5 — sub-100ms full-text search on-device because field connectivity is unreliable. A Pydantic-validated ingestion pipeline ate every paper catalog, every supplier PDF, and every legacy spreadsheet. The result: 100% normalized digital catalog coverage for the first time in the company's history.
New catalog pages still come in on paper. The same mobile app scans them, and they land in a review queue instead of a binder.
Lookup time dropped from 15 minutes to 30 seconds per job. That's a 30× speed factor, but the number that matters commercially is daily throughput: technicians now close five jobs per day instead of three. The business added 67% more service capacity without adding a single new hire — which means the margin on the growth is close to unit economics, not diluted by headcount.
Dispatchers get a second benefit: live visibility into which parts are being pulled where. The system has already caught two instances of stock imbalance before they caused a second drive.
Any field-service operation drowning in paper documentation — HVAC fleets, elevator service, agricultural equipment, marine engines — has the same shape. The combination of normalized catalog + offline-capable search + vision-model fallback for the long tail works anywhere parts identification is the rate-limiting step on throughput. Plan on 8-10 weeks from kickoff to production for a catalog of this size.
What shipped, in figures. 4 metrics.