Object detection
Bounding-box and key-point detection with multi-object tracking across frames.
AI engineering
not AI demos
Six disciplines, one accountable team. Data engineering feeds classical machine learning, deep learning carries the workloads that won't fit hand-crafted features, computer vision and NLP put models against real inputs, and MLOps keeps the system honest in production.
DISCIPLINE MAP
Six domains, opened one by one
Each discipline carries its own playbook, its own engineers and its own production history. Jump in at any layer; the team speaks across all of them.
01 Page live
Data Engineering & Preparation
The work that decides whether a model is worth training at all.
02 Page live
Machine Learning
Classical supervised and unsupervised modelling, still the right tool for most real problems.
03 Page live
Deep Learning
Neural architectures from ANN to Transformers. When the pattern is too dense for hand-crafted features.
04 Page live
Computer Vision
Pixels to decisions: detection, segmentation, OCR, medical imaging.
05 Page live
Natural Language Processing
Text and speech as first-class inputs — classification, embeddings, assistants, speech I/O.
06 Page live
MLOps & Governance
Deployment, drift, edge, bias and explainability. Where the system earns trust or loses it.
DATA ENGINEERING
Data, five stages before the first model
No model outperforms its data. We treat collection, cleaning, labelling, feature engineering and warehousing as engineering disciplines in their own right, versioned with the training code they feed.
01
Collect
Scrapers, partner APIs, streaming ingest, batch ETL. The data that reaches the warehouse is the data that trains everything downstream.
02
Clean
Outlier detection, missing-value imputation, unit harmonisation, de-duplication. Model quality ceilings are set here.
03
Label
Annotation pipelines with human review, inter-annotator agreement, golden sets. Labels that survive the audit.
04
Engineer
Feature stores, selection under leakage constraints, transformations versioned with training code.
05
Warehouse
Lakehouse design, governance tables, retention rules, residency. The layer that stays when models come and go.
Data foundations
Open data engineering ↗ For pipelines, lakehouses and the long-lived data layer behind AI systems, the dedicated discipline page has the full scope.
MACHINE LEARNING
Supervised, unsupervised, and the discipline around them
For most business problems the right answer is a well-built gradient-boosted model with clean features, not the largest transformer in the lab. We still do the transformer work; we just don't start there.
Supervised
Learn from labelled outcomes
- Linear · logistic regression
- Gradient boosting (XGBoost · LightGBM · CatBoost)
- Random forests
- SVM · kernel methods
- Neural classifiers
Unsupervised
Structure without labels
- K-means · hierarchical
- DBSCAN · HDBSCAN
- PCA · t-SNE · UMAP
- Gaussian mixture
- Autoencoder embeddings
Discipline
How we keep the model honest
- Train · validation · test partitioning
- Stratified k-fold cross-validation
- Hyper-parameter optimisation (Optuna · grid · Bayesian)
- Live A/B testing with guard metrics
- Forecasting (ARIMA · Prophet · N-BEATS)
DEEP LEARNING
A stack of architectures, each with its own reason
We pick architectures on task shape, not vintage. Feed-forward nets still land; convolution still owns images at scale; recurrent state still beats attention on some signals; transformers dominate the rest. Fine-tuning is the default before full training.
1960s+ ANN
Feed-forward networks
Dense layers, back-prop. Still the simplest baseline when tabular structure is lost.
1990s+ CNN
Convolutional networks
Translation-invariant feature detectors for images, spectrograms and structured 2D signals.
1997+ RNN · LSTM
Sequential networks
State across time for text, audio and telemetry. Superseded but not obsolete.
2017+ Transformer
Attention architectures
Self-attention over tokens. The substrate for modern language, vision and multi-modal models.
2014+ Generative
GAN · Diffusion · VAE
Distributions learnt well enough to sample from. Images, speech, molecules, code.
now Transfer
Fine-tune · adapter · LoRA
Pretrained weights plus targeted training. The pragmatic default when budget and data are finite.
COMPUTER VISION
From pixels to decisions
Vision systems that go to production need more than a strong backbone. Dataset design, calibration, latency budgets and domain-specific review loops are where the release stands or falls.
Segmentation
Semantic and instance segmentation for autonomy, inspection and medical domains.
Face & emotion
Recognition, verification and affect analysis under privacy-aware configuration.
Image enhancement
Denoising, super-resolution, deblur, restoration for low-signal input.
OCR & document AI
Layout parsing, table extraction, handwriting recognition, structured document output.
Medical imaging
MRI, CT, X-ray, histopathology pipelines with radiologist-in-the-loop review.
Deeper vision work
Open computer vision ↗ Dataset design, calibration, inspection pipelines and medical-imaging review loops live on the dedicated computer vision page.
NATURAL LANGUAGE
Text and speech, built to hold up under real use
Most NLP "wins" collapse on messy, multilingual, domain-specific input. We tune encoders on in-domain data, pair LLMs with retrieval instead of trusting open-book inference, and ship speech with diarisation and noise guards.
Sentiment & classification
Review mining, intent detection, moderation queues — fine-tuned encoders outperform zero-shot at scale.
Embeddings & vector search
Word, sentence and domain-tuned embeddings. The retrieval layer underneath modern assistants.
Question answering & assistants
Closed-book LLM, retrieval-grounded QA, multi-turn dialogue with tool use and guardrails.
Translation & summarisation
Domain-tuned MT, long-form abstractive summarisation, style-preserving rewriting.
Speech I/O (STT · TTS)
Streaming transcription, speaker diarisation, neural voice synthesis for product and accessibility surfaces.
OPERATIONS
Where the system earns trust, or loses it
Deployment is the easy half. Monitoring, drift detection, on-device compression, bias audit, privacy engineering and explainability are the layer that turns a demo into an accountable production system.
Deployment & runtime
- REST / gRPC services (FastAPI, Flask, BentoML)
- Autoscaling GPU inference (vLLM, TGI, Triton)
- Batch + streaming prediction paths
- Blue-green and canary rollouts
Monitoring & iteration
- Performance and latency SLOs
- Data drift and concept drift detectors
- Shadow traffic eval, retraining triggers
- Incident runbooks with on-call ownership
Optimisation & edge
- Quantisation (INT8, 4-bit)
- Pruning and distillation
- CPU · ARM · mobile compilation (ONNX, CoreML, TFLite)
- On-device inference with sync contracts
Governance & trust
- Bias measurement (demographic parity, equalised odds)
- Privacy: anonymisation, differential privacy, data minimisation
- Explainability (SHAP, integrated gradients, attention probes)
- Audit trails and model cards
BRANCHES
Where each discipline continues
Three pages live today; three are scoped and on the near roadmap. Each carries its own playbook, partners and production history.
Branch 01 · Live
Data Engineering
Pipelines, lakehouses and the data layer that makes AI reproducible and compliant.
Open page → Branch 02 · LiveMachine Learning Engineering
Classical ML and deep learning, training pipelines, evaluation suites, model registries.
Open page → Branch 03 · LiveComputer Vision
Detection, segmentation, OCR, medical imaging and video analytics pipelines.
Open page → Branch 04 · LiveDeep Learning
Architecture selection, distributed training, generative systems, fine-tuning and inference optimisation.
Open page → Branch 05 · LiveNatural Language Processing
Domain-tuned encoders, retrieval-grounded assistants, translation, speech I/O and safety engineering.
Open page → Branch 06 · LiveMLOps & Governance
Deployment, drift detection, inference optimisation, bias audit, XAI and privacy engineering.
Open page → Discipline · not demo
From first dataset to live inference, one accountable team
Bring a problem. We come back with a feasibility note, cost envelope and first-stage plan inside ten working days. No discovery theatre, no capability decks, no model-name bingo.