Pre-built APIs hit 93% on clean demos and 45% on your actual handwriting. Cloud vision models look great in the lab and fall apart on your factory floor's real lighting. We build NLP and computer vision systems that close that gap: trained on your data, calibrated to your environment, with the rules and exception handling that production needs.
Pixelfield is for CTOs, VPs of Engineering and operations leads who've already tried the API, hit the wall and need production accuracy on real-world documents, images and text. The seniors who scope the work also write the code. 50+ AI features in production, fastest deployment two months from kickoff to live, and a reputation for telling you when Vision API is enough.
Trained on your data, not benchmarksHybrid: model + rules + exception handlingEnvironmental calibration for CVHonest about when pre-built wins
Models score great on benchmarks and fall apart on real data. Your first production runs typically land at 25-50% of what you saw in the lab and need ten times the effort to recover. We design the gap into the scope from week one: paid PoC on your real data, accuracy reported by data type (printed vs handwritten, lab vs floor) and explicit failure-mode analysis.
Cloud OCR APIs sell 95% on clean print. On real handwriting they hit 45-50%. NLP APIs hit 80-90% on standard text and drop fast on your industry jargon, slang and multilingual edge cases. We close those gaps with specialised models, fine-tuning on your labelled data, and the post-processing rules every production system needs.
A model that returns perfectly formatted JSON with the wrong values is the worst kind of failure. We design confidence calibration, exception handling and human review into the workflow: low-confidence outputs route to review, layout shifts trigger alerts, drift on a sliced metric pages an engineer. Silent failures are the ones that destroy trust.
Production NLP and CV is rarely just a model. It's model plus rules engine plus post-processing plus exception handling plus human review plus feedback loops. One real NER pipeline ran with 'a million additional rules on top to fix the results'. We build the whole stack and explain why each layer earns its place.
Are you looking for these services?
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.
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.
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.
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.
The questions CTOs and operations leads ask us in the first call about <strong>NLP, computer vision, accuracy on real data and production cost</strong>.
Often dramatically. Cloud OCR APIs hit 93-95% on clean print and 45-50% on real handwriting. Specialised models reach 95% on the same handwriting. CV models trained on COCO collapse on your factory floor due to background leakage and lighting; site-specific fine-tuning on roughly 200 real images usually fixes it. We test the API on your data in discovery so you know exactly which gap you're closing before you commit.
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.
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 95%+ accuracy. We'll tell you which side you're on before quoting a build.
Yes, but rarely on cloud APIs. One reference Rekognition deployment hit around $2,280 per camera per month. We deploy on-device using ONNX Runtime, TensorRT Lite, TFLite or Rust wrappers, with hardware-validated accuracy. The same INT8 model can drift between 71% and 93% across different Snapdragon chips, so we test on the silicon you actually plan to ship on.
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.
Discovery is fixed-fee from £2K. Production builds typically start at £10K 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 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.
