AI / ML Models

The Models Behind Our Products

Every intelligent feature in our software is powered by a deliberate model choice — here's the full picture.

Home Models
12+
Models in Production
95%
Avg. Model Accuracy
2M+
Inference Calls / Month
5
AI Domains Covered

Five Domains of Intelligence

Our AI/ML work spans five distinct capability areas, each serving a different product need.

Natural Language Processing

Text understanding, generation & classification

Used in chatbots, document summarisation, sentiment analysis, and automated report generation across our healthcare and education platforms.

BERT / RoBERTa GPT-4o API spaCy NER Sentence Transformers

Predictive Analytics

Forecasting, anomaly detection & risk scoring

Powers demand forecasting in procurement, student performance prediction in education platforms, and financial risk scoring for government disbursements.

XGBoost LightGBM Facebook Prophet Isolation Forest

Computer Vision

Image classification, OCR & object detection

Automating document verification, ID scanning, and visual inspection workflows. Used in healthcare onboarding and government identity pipelines.

YOLOv8 EasyOCR OpenCV ResNet-50

Recommendation Systems

Personalisation, ranking & relevance

Personalised learning pathways in CampusCare360, supplier ranking in Procurit, and content relevance scoring for internal knowledge bases.

Collaborative Filtering Matrix Factorisation FAISS

Generative AI

Content creation, code assistance & synthesis

Automated report writing, proposal generation, and internal code review assistants. Built on top of leading foundation models with RAG for domain grounding.

GPT-4o Gemini 1.5 RAG Pipelines LangChain

Graph & Relational Learning

Network analysis, fraud detection & knowledge graphs

Detecting anomalous procurement patterns, mapping organisational relationships, and building knowledge graphs for internal search and discovery tools.

Graph Neural Nets NetworkX Neo4j

How We Choose the Right Model

There's no single best model — the right choice depends on the task, the data, and the constraints.

01

Define the Task Precisely

Classification vs. generation vs. retrieval vs. ranking. Before picking a model, we write an explicit problem statement with measurable success criteria — accuracy, latency, or cost targets.

02

Start with the Simplest Baseline

A logistic regression or a TF-IDF search often outperforms a large model on small, structured datasets. We always benchmark a simple baseline before adding complexity.

03

Evaluate on Real Data

Model benchmarks are marketing. We evaluate on a held-out slice of the actual production data distribution — including edge cases and adversarial examples.

04

Consider Total Cost of Ownership

Inference cost, fine-tuning cost, maintenance burden, and vendor lock-in all matter. A slightly less accurate open-source model that runs on your own infra may be the right call.

From Experiment to Production

A model that only works in a notebook is not a product. Our MLOps pipeline takes experiments to production reliably and repeatably.

Data Pipeline

Raw data ingestion, cleaning, feature engineering, and versioning with DVC or Delta Lake.

Experiment Tracking

Every training run logged in MLflow — hyperparameters, metrics, artefacts, and environment.

Model Registry

Versioned model artefacts with staged promotion: dev → staging → production.

Serving Layer

REST APIs via FastAPI or TorchServe, with autoscaling and circuit breakers.

Monitoring & Drift

Production data distribution tracked continuously. Drift triggers automated retraining pipelines.

Ethical AI Principles

Every model we deploy goes through a responsible-use review. These are not aspirational values — they are checklist items.

Explainability Decisions that affect users must be explainable. We use SHAP, LIME, or attention visualisations where required.
Fairness Audits Models are tested for disparate impact across demographic groups before deployment.
Data Privacy PII is anonymised before training. Models are evaluated for potential data leakage.
Human-in-the-Loop High-stakes decisions (medical, financial) always have a human review step before action.
Model Rollback Every deployed model can be instantly rolled back to the prior version if degradation is detected.

Want AI Built Into Your Product?

We'll assess your data, define the right model strategy, and build production-ready AI features from scratch.

Start a Conversation See Our Full Tech Stack