Two bodies of work. One that companies paid for and other people now depend on, and one I wrote alone because I wanted to know how the parts fit. The second is why I am useful on the first.
Production work across four industries. Employer systems are described without internal product names.
2024 - present · team of 5
An internal multi-agent chatbot that connects Sales and Marketing teams to their own data warehouses in natural language. Tool-augmented agents plan, call SQL and Python, reason over warehouse-scale data, return charts and explain how they got there. I lead the team of five that builds it.
2024 - present
A sales-intelligence platform that pulls third-party signals on enterprise accounts, vectorises them, and serves a multi-agent reasoning layer that proposes outreach plays to the field.
2022 - 2024
A talking-head conversational avatar, built end to end: Stable Diffusion XL for visuals, GFP-GAN for face restoration, BARK for voice, and Llama 2 with LangChain and PEFT/LoRA for the conversational core and its personality fine-tunes.
2022 - 2024
A voice assistant shipped on Skullcandy's Skull-iQ headphone platform. Rasa NLU and Dialogflow handled intent, Flask served inference, GKE handled scale.
2022 - 2024
A UNet2D + DDPM diffusion model written from scratch on a custom domain dataset - built to internalise how diffusion actually works rather than calling a hosted API and hoping.
2020 - 2022
Categorisation models over an ecommerce catalogue of millions of SKUs - hybrid CNN + RNN and VGG16 + BiLSTM architectures - plus YOLOv5 detectors for fine-grained part detection on manufacturing floors.
IoT
An on-device computer-vision pipeline for a connected refrigerator - inventory tracking, freshness estimation and anomaly alerts, deployed on Azure ML.
2022 - 2024
Led the migration of the team's Dataiku workflows onto a proper CI/CD pipeline. Team productivity up 20% - the one number here I can put a figure against.
Written on my own time, on one 6 GB laptop GPU, mostly to find out where the abstractions leak.
2025 - present
A deep reinforcement learning trading system covering 1,400+ stocks on a custom multi-factor environment with DQN agents. Feature engineering across price, volume and indicator history; reward shaping; and a backtest loop I trust enough to argue with. It is private because the codebase cannot leave my machine - which is exactly the constraint that pushed the rest of my work toward local inference.
2025
An agentic text-to-SQL and analysis platform. It runs RAG over hand-written schema documentation (tables, columns and join paths) so the model generates SQL against structure it actually understands, registers SQL / Python / visualisation as tools in a single LLM call, and caches results in session memory so follow-up questions reuse data instead of re-querying. Generated Python does the statistics; Plotly does the charts. Local throughout, on Ollama with qwen3:8b.
2026
Hooking a fully local agent up to my own backtesting output so I can ask questions in plain language - which stocks earned the best rewards, where the policy degraded - without shipping proprietary analysis to anyone's server. A running record of what an offline agent loop costs in latency and capability, and what it buys back in control.
2020
A recurrent neural network classifier for spam, built in TensorFlow 2 - one of the first models I took all the way from raw text to a working notebook, and still a clean reference for the preprocessing-to-evaluation loop.
2023
Pipeline scaffolding for training and deploying models on Azure ML - the plumbing between a notebook that works and a model that other people can actually call.
Not my work - repositories I keep forked and open while I figure out how their authors solved something. Listed because what you read shapes what you build.
Public repositories, forks and the occasional experiment that did not survive.