Most companies sit on data they're not using and make manual decisions a model could automate, audit and improve. We build classical ML systems for prediction, classification, anomaly detection, forecasting and recommendations that ship to production and stay accurate.
Pixelfield is for CTOs, data leads and operations heads at companies where the obvious answer to your problem is a model, not a chatbot. 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 XGBoost beats a £40,000 LLM build.
Classical ML where it beats GenAIBusiness metrics, not just accuracyProduction-first, not notebook experimentsDrift monitoring built in
In 2026 the default ask is 'use GPT for everything'. For structured tabular data with labels, classical ML is faster to train, cheaper at scale (CPU pennies vs LLM tokens), more deterministic and easier to audit. We make the call with you in writing during discovery: classical for prediction on tabular data, GenAI when the input is genuinely unstructured. Around 80% of structured business problems land on the classical side.
A churn model with 99% accuracy on imbalanced data is useless. A retention campaign with 92% false positives can lose more revenue than it saves. We define the business-impact metric (net revenue, cost saved, decisions automated) before we choose the model, then measure against it through shadow testing and A/B evaluation. AUROC is a sanity check, not a goal.
Across the industry, around 80% of ML models never ship past the notebook. We treat the model as the easy 20%. The other 80% is data pipelines, feature engineering, model serving, drift detection, retraining, observability and the cross-functional sign-off needed for Legal, Operations and Finance to put it live. We design that scope from week one.
Logistic regression on six months of engineered features beats deep learning on raw data more often than the conference talks let on. We invest in understanding your data first: schemas, distributions, leakage, lineage, label quality. Then we pick the simplest model that solves the problem at the latency and cost you can afford. Sometimes the answer is a rules engine. We'll say so.
Are you looking for these services?
We map your problem against your data: volume, label quality, feature availability at prediction time, leakage risk, lineage. Discovery includes a paid PoC on your real data when feasibility risk is high. We test the simplest model first to set a baseline.
Output: feasibility report, classical ML vs GenAI recommendation, projected business impact, fixed-scope build quote, or an honest 'your data isn't ready yet, fix this first' if that's the answer.
Most of the work lives here. Feature design, leakage checks, train-validate-test splits that respect time and identity, model selection driven by your latency, cost and explainability constraints. Build runs as a small senior team led by Michal Vavra, embedded with your data and operations stakeholders.
Each iteration is gated on a fixed eval set with business-metric guards, not vanity accuracy.
Before the model drives any decision, we run it in shadow mode against your real production traffic for two to twelve weeks, comparing predictions to actual outcomes. This gives a true accuracy and ROI projection before you spend a penny on interventions. Then we ship behind feature flags with a manual override and rollback path.
We've stopped builds at this stage when the economics didn't justify production. That's what shadow testing is for.
At launch we wire in prediction-confidence histograms, business-outcome monitoring, drift detection, retraining triggers and alerting. Standard latency dashboards do not catch the model returning perfect 200s while drifting silently. We monitor what actually moves: business metrics first, then statistical drift.
Monthly retainer for monitoring, retraining and iteration, not a lock-in. Take the system in-house whenever you're ready.
The questions data leads and CTOs ask us in the first call about <strong>classical ML, GenAI, drift and run-rate cost</strong>.
If your data is structured and labelled (transactions, customer records, sensor data, time series) and you need prediction, classification or anomaly detection, classical ML usually wins on accuracy, cost, latency and explainability. CPU-only inference at pennies per million predictions, deterministic outputs, full feature importance. LLMs win when the input is genuinely unstructured (free-text feedback, documents, multi-modal). We make the call with you in writing during discovery, not after you've spent £40K on prompt engineering.
Around 80% of ML models stall in the notebook. The model itself is the easy 20%. The 80% that kills projects is data quality, feature pipelines, model serving, drift detection, retraining, observability and the cross-functional sign-off needed for Legal, Operations and Finance to put it live. We design that scope from week one. The model that ships is the one we've already wired into the production pipeline before we tune the hyperparameters.
It depends on the problem. For most classification and regression on structured data, a few thousand to tens of thousands of labelled examples. Anomaly detection can work with fewer examples of normal behaviour. Time-series forecasting needs enough history to cover the seasonality you care about. We assess your data quality, volume and label coverage in discovery and tell you honestly when the answer is 'fix the data first, then revisit'. We've recommended that more than once.
The business outcome, not the AUROC. We define the impact metric (net revenue, cost saved, decisions automated, false-positive cost) before we pick the model. A churn model that raises retention but loses net revenue is a failure. A fraud model with 99% accuracy on imbalanced data is meaningless. Accuracy, precision and recall are sanity checks. The metric we report on is the one your CFO would recognise.
Yes, models degrade as your data and your customers change. Standard monitoring (latency, error rates, GPU utilisation) does not catch it. We ship every model with prediction-confidence histograms, business-outcome monitoring, drift detection on feature distributions, golden-dataset regression and retraining triggers. Retraining cadence depends on volatility (weekly for fraud, monthly for tabular classification, on-trigger for everything else). Maintenance retainer covers the work, or we hand it over.
Discovery is fixed-fee from £2K. Production builds typically start at £10K and run six to ten weeks. Run rate is a fraction of GenAI: classical ML on CPU costs pennies per million predictions, plus a flat retainer for monitoring and retraining. We quote both numbers in the proposal so the run rate is clear before you commit.
When a rules engine, a SQL query or a simple business heuristic does the job. When your data is too messy, too sparse, or doesn't predict the outcome you care about. When the cost of false positives outweighs the value of correct predictions (we've seen 92% false-positive rates on real churn projects). We'll say so during discovery and point you at the cheaper, more reliable solution. We'd rather not build it than build the wrong thing.
