Service · Computer vision

Read what's in the image, the video, and the scanned PDF.

Vision-language models in 2026 are good enough that 'extract X from this image' is no longer a research project. We build production pipelines for document understanding, product imagery, manufacturing QA, and logistics - using VLMs when they fit, classical computer vision when it's cheaper, and both when the problem deserves it.

Last updated:

The problem

Hours per day spent reading what nobody has time to read.

A meaningful slice of business operations runs on images and PDFs nobody has time to read carefully - invoices, contracts, product photos, shipment manifests, security footage, customer-uploaded receipts. The work eats hours per day; outsourcing it adds latency without solving the cost long-term.

Off-the-shelf OCR catches the easy cases. Modern VLMs catch the rest, if you build the pipeline carefully - with verification, edge-case handling, and clear escalation when the model is not sure.

What's included

Six things every vision pipeline gets.

  • Discovery

    Sample your real images / documents / video. Identify the variation that breaks naive approaches.

  • Pipeline design

    VLM-only, classical CV, or hybrid - which combination fits your cost, latency, and accuracy target.

  • Implementation

    Typed extraction schema, confidence scoring, human-in-the-loop for low-confidence outputs.

  • Validation harness

    Golden set of 50-500 examples; regression testing on every model update.

  • Production deployment

    Async queue, monitoring, audit log of every decision.

  • Monthly review

    What edge cases are coming through, where to tune, when the model needs a refresh.

How it works under the hood

Why pure-OCR and pure-VLM both fail - and the hybrid that ships.

Either extreme has a recognisable failure mode. The hybrid pipeline is what survives in production.

Where pure-OCR or pure-VLM pipelines stall
common patterns we have watched up close
  • OCR-only - brittle on layout variation, fails on handwriting, mixed scripts, low-quality scans
  • OCR-only - no semantic understanding, no reasoning about what the document means
  • VLM-only - too expensive at production volume; cost-per-page does not scale
  • VLM-only - confidence is hidden inside the answer, hard to gate on
  • No audit trail; impossible to investigate when an extraction goes wrong months later
How we build the hybrid
classical CV first · VLM on edge cases · human review on low confidence
  • Classical CV handles the high-volume / easy cases at low cost
  • VLM handles the edge cases - rare, expensive, but bounded
  • Confidence score per output gates human review
  • Cost stays bounded as volume scales
  • Audit log + confidence score per decision - debuggable months later
Use cases we ship

Real vision pipelines in production.

  • Invoice + receipt extraction

    Line items, totals, VAT, party details. Match against POs. Flag anomalies. Books close faster.

  • Contract intelligence

    Pull dates, parties, obligations, renewal terms from PDFs in any layout. Speeds legal review by 3-5×.

  • Product photo QA

    Flag mislabelled SKUs, check that the photo matches the catalogue description, generate alt text + descriptions.

  • Manufacturing QA

    Visual defect detection on production-line photos. Often a hybrid VLM + classical pipeline.

  • Logistics manifests

    Read shipment paperwork, match against ERP records, flag discrepancies before goods leave the warehouse.

  • Security / monitoring

    Video summarisation, event detection, alert generation from a monitored feed.

Case study

Every frame of 119 TB of footage, labelled - and searchable by plain-English prompt.

A global events company had a decade of footage - keynotes, interviews, drone passes, b-roll - that nobody could search through. We fine-tuned a vision-language model on the agency's recurring speakers, stage layouts, and branding so it recognised the people and places that mattered. Then we processed the 119 TB archive at 8 frames per second sustained throughput - about 28 days of continuous processing - and built a plain-English search interface on top of the per-frame metadata. Crowd footage stayed inside a strict GDPR boundary: identity is only attached to a frame when the person is on the consent register; everything else is searchable by clothing, posture, scenery, on-screen text, and dominant palette - never by face. New uploads go from drop to fully indexed in under an hour.

119 TB
Historical footage indexed frame-by-frame
~28 days
Continuous historical processing pass
8 fps
Pipeline throughput sustained
<1 hour
Upload-to-indexed for new event footage
FAQ

Three things people always ask.

What about handwriting?
Modern VLMs are good at handwriting in well-represented Latin scripts. Cyrillic, Greek, Arabic, and other less-represented scripts are harder; we test on a sample of your real documents before committing to a accuracy target.
What about live video?
Yes. We've shipped a few real-time pipelines (mostly security / monitoring). Cost-per-second-of-video adds up fast - we are conservative about when this is the right tool.
Can it run on our hardware?
Yes - DGX Spark or your hardware. Particularly important when imagery is sensitive (medical, security, legal).
AI that already runs

Bring us the documents nobody has time to read.

30-min diagnostic call walks through what you process today and what a hybrid pipeline would cost. Free; written summary either way.

Live · powered by Gemma-4 · running on our hardware in Sofia