Your Data. Your Problem. Your Model. Engineered From First Principles.
Custom model development is the discipline of building bespoke predictive or perceptual models on an organisation's own data, against an organisation's own problem, under an organisation's own constraints. It is the opposite of dressing up a pre-trained API with a thin layer of configuration. It is the opposite of grid-searching XGBoost on a cleaned-up spreadsheet and calling it a model. Done properly, a custom model compounds into a proprietary advantage that competitors cannot replicate by buying the same tools. Done poorly, it becomes a notebook that nobody trusts, a dashboard that nobody opens, and a line item that nobody defends in the next budget cycle. Entiovi's Mintaka practice builds the first kind.
A custom model is not the right answer for every predictive problem. It is the right answer in specific, measurable situations — and knowing the difference is the first piece of engineering.
Custom model development earns its place when an organisation has proprietary data that general-purpose models cannot see — customer behaviour, operational telemetry, claims history, transaction streams, inspection images, call recordings. It earns its place when the target is a specific business metric — an approval rate, a cycle time, a loss ratio, a service level — and not a generic score that has to be reinterpreted. It earns its place when accuracy, calibration, or latency requirements exceed what off-the-shelf offerings can deliver. And it earns its place when regulatory or audit constraints demand full documentation, reproducibility, and control over the training data, the weights, and the deployment.
A tuned API will not learn a firm's customers, products, or risk posture. A bespoke model will — provided it is engineered with discipline. That qualifier is the entire point.
Custom means several specific things in Mintaka engagements, and none of them are loose.
The training corpus is the client's own, under the client's own licences, residency rules, and retention policies. Data never crosses a boundary that has not been explicitly sanctioned. Where data must remain on-premise or inside a sovereign region, the training runs there.
The decision the model informs, the metric that moves if it works, and the cost of a wrong answer are agreed before any training begins. Models built against a vague objective produce vague results.
Every custom model is scored against the incumbent — a rules engine, a vendor score, a statistical baseline, or a heuristic — on an honest live hold-out, not only against academic benchmarks. A lift number with no baseline is a marketing line, not an evaluation.
Model card, data sheet, training log, evaluation pack, risk statement, and reproducibility manifest ship with every production model. The documentation is the audit evidence.
Artefacts, weights, code, pipelines, and the right to retrain remain with the client. No lock-in, no hidden runtime dependencies, no rented inference.
Model selection is a function of the data, the decision, and the deployment envelope — not of preference or familiarity. Mintaka engagements span the full family spectrum.
Gradient-boosted trees (XGBoost, LightGBM, CatBoost), generalised linear models, support vector machines, calibrated classifiers, and stacked ensembles. Still the workhorse for tabular prediction and decisioning at enterprise scale.
Multilayer perceptrons, convolutional networks, Transformers, and recurrent architectures (LSTM, GRU, TCN) applied to structured, text, image, audio, and sequence data — where the signal is rich enough to reward representation learning.
Clustering, anomaly detection, representation learning, and dimensionality reduction for segmentation, portfolio structuring, outlier discovery, and exploratory analysis.
Where labelled data is scarce and the cost of labelling is high — industrial inspection, rare-event detection, specialist domains — pretraining on unlabelled data and fine-tuning on a smaller labelled set is frequently the decisive choice.
Graph Neural Networks for relationship-rich domains — fraud networks, recommendation systems, knowledge graphs, supply-chain topology — where the signal lives in connections as much as in features.
Classical baselines with deep-learned features, ML models wrapped around business rules, or deep encoders feeding calibrated classical heads. Portfolio-level performance almost always beats single-family purity.
Every model Mintaka ships passes through six engineered stages, each with a gate, a deliverable, and a reason to stop if the gate fails.
What decision will this model inform? What metric moves if it works? What is the cost of a wrong answer, and is that cost symmetric? The answers define the evaluation plan before any modelling begins.
Availability, quality, bias, lineage, leakage, fairness, and target-contamination risk — examined and documented before training. Most model failures are data failures wearing a model's clothes.
Transformations, aggregations, encodings, point-in-time correctness, and representation learning where it earns its place. The feature layer is versioned and tied to the model registry from day one.
Reproducible, versioned runs across candidate families. Every experiment tied to a data version, a code commit, and an evaluation report. Bayesian and Hyperband search where the cost of search justifies it — not notebook grid loops.
Holdout, cross-fold, time-split, and backtest validation chosen for the data shape. Calibration of probabilistic outputs and decision thresholds to the business loss function. The model is not finished when accuracy is satisfied — only when calibration is too.
Container, model card, inference contract, training pipeline, retraining plan, and runbook delivered as a single, deployable asset. Handover includes the training required for the client team to operate it independently.
A model that is accurate but miscalibrated is dangerous in a decisioning system. It over- or under-states confidence, and downstream logic acts on the miscalibrated probability. Mintaka treats calibration with the same rigour as accuracy.
Expected calibration error, reliability diagrams, cost-sensitive thresholds, precision-recall at the operating point, and time-sliced performance are all standard evaluation outputs. The threshold at which the model flips from one decision to another is tuned against the asymmetry of a false positive and a false negative — because in most enterprise problems, those costs are not equal.
Models ship only when accuracy and calibration are both satisfied. Never one without the other.
Every production model Mintaka ships can be defended — technically, commercially, and where relevant, legally.
Explainability artefacts are chosen for the model family and the audience. SHAP, LIME, integrated gradients, and counterfactuals for deep and ensemble models; coefficient tables and calibration curves for linear and generalised linear models; attention and saliency for vision and sequence architectures. The audience determines the depth — a data scientist does not need the same artefact as a risk reviewer, a regulator, or an end user.
Fairness is evaluated across protected attributes where relevant, with documented decisions where trade-offs exist. Model cards, data sheets, training logs, and reproducibility manifests are produced as part of the build — not retrofitted to satisfy a validation team after go-live.
Custom model development engagements cluster in domains where proprietary data compounds into durable advantage.
Propensity, churn, lifetime value, and next-best-action models for customer operations.
Credit risk, fraud detection, and portfolio scoring for financial services.
Demand elasticity, pricing, and promotion-response modelling for retail and consumer goods.
Claims severity, frequency, and reserve prediction for insurance.
Preventive maintenance and equipment failure prediction for industrial operations.
Quality inspection and defect detection models for manufacturing and assembly.
Recommendation and ranking systems for customer-facing surfaces.
Problem frame, data landscape, success metric, cost-of-error model, deployment envelope.
Feature strategy, candidate model families, evaluation protocol, calibration plan.
Pipelines, training, validation, versioned artefacts, experiment log.
Thresholds, probabilistic outputs, fairness review, explainability pack.
Inference contract, model card, data sheet, risk documentation.
Repository, runbook, retraining plan, monitoring hooks, client-team training.