Stop doing

manually.

UK agencies and B2B firms use me to automate the workflows their teams manually slog through every week, custom-built to exactly how you work, and running within days.

What I build

Live demos

https://outreach.ishsitotombe.co.uk

Enter your name or business name and I'll generate a personalised cold email, sent straight to your inbox.

Sales · CRM · Companies House

Outreach Agent

Search UK companies via Companies House, enrich with director data, generate personalised cold emails, then manage the full pipeline. Give it your details and see a real email land in your inbox.

Open source

Built in public.

Swift · macOS · AI Agent

MCP-x-Mac

Gives AI agents direct control of any Mac app, including ones with no API, through a self-building tool registry.

A self-evolving MCP server that discovers its own capabilities, writes its own tools, and self-heals when apps update. 71 tools across 50+ apps. 80 tests.

Self-healing71 tools50+ appsMIT open source
HuggingFace · Dataset · LLM Training

Advanced Math Error Correction

A training dataset that teaches AI models to catch and fix their own reasoning errors, the same rigour behind the automations I build.

Curated dataset for fine-tuning LLMs on mathematical reasoning and error correction. Designed to improve model accuracy on multi-step problems where small mistakes compound.

Fine-tuningMath reasoningOpen dataset
Research · Continual learning

Testing a slime-mould idea against Fisher information

A pre-registered continual-learning study: my hypothesis was wrong, and the way it was wrong was measurable.

A biology-inspired importance signal beat the standard method at preventing catastrophic forgetting on an easy benchmark, then failed on a harder one. I tracked down exactly why.

Pre-registeredFalsified + explainedRead the write-up
Research · Multi-agent AI · Local inference

HSI: Hierarchical Sparse Intelligence

22 specialist classifiers pre-process every query before an LLM ever sees it. The architecture, the benchmarks, and where it falls short.

A proof-of-concept multi-agent pipeline: fine-tuned DeBERTa classifiers compress language, syntax, intent, math and logic into a structured buffer that the LLM reads instead of raw text. Runs fully local via Ollama.

22 specialistsLocal-first (Ollama)MIT open source
How it works

Three steps, then it runs itself.

01

Tell me what's eating your time

Click 'Get in touch', takes 2 minutes. I'll reply within 24 hours with a quick call to understand the problem.

02

I build the automation

I build with the best AI tooling available, which means faster turnaround and a fraction of what a traditional dev shop charges.

03

You get your time back

The automation runs in the background. Most projects are a one-off cost, no monthly fees.