Agentic systems, LLMs, computer vision, and applied game-AI.
Modern decoder-only transformer (RoPE, RMSNorm, SwiGLU, flash attention) built from scratch as an optimisation sandbox.
A foundation for research into small, locally-hosted reasoning models - architectures and training recipes you can take apart and modify, instead of treating an API as a black box.
Extendable digital analyst - answers natural-language questions over a marketing graph with charts, profiles, and evidence-backed recommendations.
A digital analyst that joins ambiguous business questions to schema-safe Cypher queries, charts, and human-readable narrative - so growth teams act on data instead of waiting on a data-science queue. The same supervisor + specialist-team pattern is portable to any domain with rich relational data.
Real-time voice agent for advisory conversations - runs fully on-device, end-to-end.
Lets regulated industries (finance, healthcare, legal) deploy conversational AI without sending customer audio to a third-party API. Airtight on data residency.
Churn-risk scoring with built-in reasoning traces - symbolic + Bayesian + LLM layers.
Gives customer-success teams a why alongside the score, not just a number. Critical when acting on a prediction requires sign-off from a human who needs to justify the call.
Full-stack editor where an LLM drafts claims, tracks dependencies, and flags inconsistencies live.
Compresses the expensive, high-precision work of patent drafting into a tight human-in-the-loop workflow. Same pattern fits any specification-heavy legal or compliance document.
One-line embeddable chat widget backed by a self-hostable RAG service grounded in your own docs.
Lets any business drop a grounded AI assistant onto their site without integration work - the same pattern that powers customer support, internal helpdesks, and product onboarding.
Reinforcement-learning generator that produces 100% solvable 2D platformer levels.
Demonstrates RL-driven content generation with guaranteed validity constraints - a pattern that transfers to any domain where generated outputs must meet hard rules.
Tiny PyTorch MLP that regresses sin(x) - the smallest end-to-end project that teaches train/val splits, MLflow tracking, and out-of-distribution failure.
A teaching scaffold I use with students - the model is a few lines, but the experiment infrastructure around it is the same shape they will use on real work.
Multi-stage pipeline - segmentation, pose, dense correspondence, garment parsing, neural refinement.
Reduces returns and friction in e-commerce by showing the shopper the garment, not a stock photo. Also a reusable CV stack for any body-aware retail or fitness product.
Vision-language agent that grounds goals into real keyboard/mouse actions across any app.
A template for operating legacy enterprise software without custom APIs. Relevant anywhere a long-tail of GUI tools blocks wider automation.
Fine-tuned generative model producing structured sprite sheets at production quality.
Shows how targeted fine-tuning + pipeline work beats generic image models for constrained, asset-producing creative tooling.
Hybrid sentiment pipeline - BERT embeddings with KMeans clustering and a fine-tuned discriminative head.
Two ways to attack the same NLP problem - unsupervised structure-finding vs supervised fine-tuning - measured on the same evaluation harness.
End-to-end credit-risk pipeline - EDA, feature engineering, modelling, and head-to-head evaluation - on peer-to-peer lending data.
Shows the full discipline of an applied ML project, not just a model fit - the parts (EDA, leakage checks, calibration, comparison) that actually decide whether a model ships.
Logistic-regression and CNN baselines for classifying mental workload levels from raw EEG signals.
A reproducible biosignal-classification pipeline - frequency-domain feature engineering, dimensionality reduction, and apples-to-apples evaluation across linear and deep baselines.
Combat AI, behaviour trees, and enemy decision-making for a story-rich action-adventure title.
Production gameplay-AI work - behaviour trees, perception, reactive combat loops - on a shipping Unreal 5 title at O-Sen Studios.