Models trained, fine-tuned and written from nothing. Three of these are open to read; the others belong to the clients who paid for them, so they are described rather than linked.
Around 10,000 lines of hand-written C: a full forward and backward pass, with a three-factor Hebbian rule running alongside backprop and compared on the same task. Experience replay, checkpointing, a vision path, and a study of ±1 binary weights for a 32x memory cut.
An agent that answers questions about live betting markets, and can be checked.
18,000 lines of Python around a LangGraph agent: subagents, market-data and web tools, a sandboxed Python REPL, an MCP server, token streaming and async checkpointing. A test suite asserts precision and holds the streaming path to a latency budget.
A fine-tuned model that scores AI answers against a written ethical standard.
A LoRA fine-tune of Mistral-7B-Instruct with PEFT (rank 8, alpha 32, 4-bit quantisation, gradient accumulation) on a hand-built instruction dataset. Served behind an API that returns a score and a written verdict per response.
Radiology inference: segmentation and classification over CT and X-ray studies.
A production inference service on nnU-Net v2, MONAI, TotalSegmentator and torchxrayvision, reading DICOM and NIfTI, with the model runtime containerised apart from the API. Documented to the standard of a Health Canada technical submission.
A coding and security agent harness in Rust, where the orchestrator and its subagents run on different models.
A fork of xAI's Grok Build (Apache-2.0) turned into an orchestrator that plans, spawns subagents and reviews their work, with the model chosen per agent. A plan/act wall and per-tool checkpoints sit between the model and the filesystem; memory lives in an embedded PostgreSQL; browser control is a tool.
Reads a photographed betting slip and files it, without anyone typing it out.
A vision model reads a photographed betting slip sent over Telegram, returns typed fields, and files the row in a spreadsheet. Perceptual hashing stops duplicates; batching keeps the calls cheap.
Tell me what the product has to do and who it is for. I'll reply with next steps, a price, or an honest no if I'm not the right person for it.
Available for new projectsNigeria · Remote · GlobalUsually replies the same day
Five leaves, not four. The fifth one does the work.