Computer vision and document AI in London

NLP here means documents, not chatbots: invoices, contracts, messy PDFs. Computer vision is the other half: photos, cameras, handwriting. Cloud APIs look sharp on a demo. On your real inputs they often are not. We build on your data, with rules for the cases the model should not guess.

This is for teams who already tried the API, hit the wall, and need accuracy on real documents and images. We will say when Vision API is enough.

BOOK 45-MIN SCOPING CALL →see our work →References available

Trained on your data, not benchmarks
Hybrid: model + rules + exception handling
Environmental calibration for CV
Honest about when pre-built wins

4.9/5 on Google

4.4/5 on Trustpilot

5.0/5 on Clutch

Shipping software since 2013. Vision and document AI on ugly real inputs.

London team. We will say when Vision API is enough.

/ Deliverables

What We Build (and the Real-World Problem It Solves)

Document Processing and OCR (Including Handwriting)

Print is usually fine on a cloud API. Handwriting, mixed-layout forms, multi-page reports, technical drawings? Cloud APIs fall over, and generic vision models start guessing past page three. We build specialised OCR pipelines that hold up on real handwriting, structured fields and narrative text, with rules and exception handling around the model.

Reference workflows in industry have processed 150,000+ handwritten pages over a year later, on their real documents, instead of an API that needed constant human correction.

OCR

HANDWRITING RECOGNITION

DOCUMENT EXTRACTION

FORMS PROCESSING

Text Classification, Extraction and Search

Named entity recognition, intent classification, ticket routing, sentiment on real customer text. Pre-trained models hit 80-90% on clean text and drop fast on your domain jargon, slang and edge cases. We fine-tune on your labelled data, ship with confidence thresholds and human review for low-confidence cases, and tell you up front whether your data needs labelling work first.

For dedicated chat or generative work, see our LLM Development and Chatbot pages.

NER

TEXT CLASSIFICATION

ENTITY EXTRACTION

SEMANTIC SEARCH

Visual Inspection and Object Detection

Defect detection on production lines, package counting in warehouses, retail shelf monitoring, safety violations on construction sites. Lab accuracy is not factory-floor accuracy. We do site-specific calibration: lighting, camera mounting, motion budget, hard-negative mining for background leakage, drift checks against new conditions.

Pilot lab does 98%. Factory floor with windows does 38% until you fix the lighting. We address the physical environment first, the model second.

VISUAL INSPECTION

OBJECT DETECTION

RETAIL CV

SAFETY MONITORING

Edge and Real-Time Vision

Camera feeds at 30fps are where cloud APIs stop being viable. Cloud vision APIs on a live camera feed get expensive fast per camera per month. We deploy on-device or edge: ONNX Runtime, TensorRT Lite, TFLite or Rust wrappers, with hardware-validated accuracy (quantised models can score differently on different chips). Privacy, latency and predictable cost.

On-prem and edge deployments for medical, manufacturing and security use cases where data residency or sub-50ms latency rules out the API.

EDGE INFERENCE

ONNX

REAL-TIME CV

ON-PREM DEPLOYMENT

When computer vision has to work on real inputs

APIs cover the happy path

Generic OCR and sentiment APIs are fine until your documents, accents, or edge cases show up. Accuracy on your distribution is the only score that matters.

Sensitive content stays controlled

Contracts, IDs, medical notes. Processing pipelines need access rules and retention policies, not a random third-party upload.

Eval on your examples, not ours

We build test sets from your real traffic so "95% accurate" means something you can defend in a meeting.

Keep the off-the-shelf tool if…

If an off-the-shelf extractor already hits the KPI, keep it. We are for stubborn accuracy and integration problems.

Are you looking for these services?

AI Consultancy→Machine Learning→Custom AI Applications→LLM Development→

How we run a vision or document-AI build

Real-Data Audit and Feasibility (weeks 1-2)

We test the obvious pre-built option (Textract, Vision API, Rekognition, spaCy, HuggingFace) on your actual data first. Real handwriting samples, real factory floor footage, real customer text in your jargon. We measure where it holds, where it breaks, and what the failure modes look like.

Output: feasibility report with real accuracy by data slice, pre-built vs custom recommendation, environmental readiness assessment for CV, fixed-scope build quote.

Domain Adaptation and Model Development (weeks 3-6+)

For NLP: labelling strategy, fine-tuning on your data, slang and jargon coverage, multilingual handling where needed. For CV: site-specific data collection, hard-negative mining for background leakage, augmentation for lighting and motion, hardware validation across the cameras and chips you'll actually deploy on.

Build runs as a small senior team led by Michal Vavra. Each iteration is gated on a fixed eval set sliced by data type and condition, not aggregate accuracy.

Hybrid System and Production Integration (weeks 5-10+)

The model is one component. We wire in rules engines, post-processing, confidence thresholds, exception queues, human-review interfaces and feedback loops. Then we ship behind feature flags with shadow testing on real production traffic before any decision is automated.

For CV deployments we run site-specific validation across every camera, lighting condition and time of day before sign-off.

Monitoring, Drift Detection, Iteration

At launch we wire in confidence-distribution monitoring, sliced accuracy by data type, environmental drift alerts, exception-queue tracking and a feedback loop from human review back into the eval set. CV especially needs site-by-site monitoring because protocol or scanner drift can drop sensitivity 10+ percentage points silently.

Monthly retainer for monitoring, retraining and iteration, not a lock-in. Take the system in-house whenever you're ready.

What computer vision and document AI costs

Cost depends on data quality, environmental complexity (for CV), labelling effort and whether the answer is pre-built, fine-tuned or specialised custom. Discovery is fixed-fee from £2,000 and produces a defensible build quote plus a projected monthly run rate before you commit. Production builds typically start around £10,000 and run six to twelve weeks to a live system. For high-volume video, edge deployment usually beats per-frame cloud API pricing inside a quarter.

Real-Data Audit and Feasibility

From £2,000. Test pre-built on your actual data, accuracy by slice, fixed-scope quote

Site Readiness and Domain Adaptation (CV)

Lighting, camera, motion calibration and on-site data collection where the deployment requires it

Production Build

Typically £10,000+, 6-12 weeks to a live system with model, rules, exception handling and monitoring

Monitoring and Iteration Retainer

Monthly engagement for drift detection, retraining and exception-queue review

Book 45-min scoping call →Request a reference call

Frequently Asked Questions

The questions CTOs and operations leads ask us in the first call about NLP, computer vision, accuracy on real data and production cost.

We tried the API and it was terrible on our data. Will custom be different?

Often dramatically. Cloud OCR is fine on clean print and falls over on the handwriting and layouts you actually have. Off-the-shelf vision models trained on public photos collapse on a factory floor because of lighting and background. We test the API on your data in discovery, then decide if a specialised model is worth it.

How accurate will it be on our specific data?

We don't quote accuracy without seeing your data. The discovery PoC on your real samples produces honest accuracy by slice (printed vs handwritten, lab vs floor, polite vs angry user text). Then we set targets the system has to hit before launch. We also report false-positive and false-negative rates separately, because aggregate accuracy hides the failure modes that actually matter.

Should we use a Vision API or build custom?

Use the API when your task is clean printed text, basic labelling, low volume, no domain-specific accuracy requirement and no real-time/edge constraint. Build custom when you have handwriting, domain objects, real-time video (camera-feed costs explode on cloud APIs), privacy/edge requirements, or you need accuracy the API cannot hit on your data. We'll tell you which side you're on before quoting a build.

What about real-time on camera feeds? Is that even possible?

Yes, but rarely on cloud APIs. Cloud vision APIs on a live camera feed get expensive fast per camera per month. We deploy on-device using ONNX Runtime, TensorRT Lite, TFLite or Rust wrappers, with hardware-validated accuracy. Quantised models can score differently on different chips, so we test on the silicon you actually plan to ship on.

How do you stop the system from being 'confident and wrong'?

Three layers. Confidence thresholds that route low-confidence outputs to a human review queue. Output validation against typed contracts (layout shifts produce structured output, just wrong, so we check structure plus content). Sliced accuracy monitoring: we track drift on the slices that matter (per-customer, per-camera, per-document-type) so silent degradation surfaces in days, not months.

How much does a computer vision or document AI project cost?

Discovery is fixed-fee from £2,000. Production builds typically start at £10,000 and run six to twelve weeks. Run rate depends on architecture: cloud inference for low-volume NLP can be pennies per million tokens; edge CV on your hardware is near-zero marginal cost; high-volume video on cloud APIs explodes fast. We quote both numbers in the proposal so the run rate is clear before you commit.

When should we NOT build a custom vision or document-AI model?

When the pre-built API does the job at acceptable accuracy and cost. When your data is too sparse to fine-tune. When the physical environment isn't fixed (CV pilots fail because of lighting, not the model). When a rules engine or template-based extraction would be more reliable. We'll say so during discovery and point you at the cheaper, more reliable option.

Related AI work

AI software development is the product work. This page is computer vision and document AI on ugly real inputs.

AI software development→AI development company→RAG development→