SELECTED WORK

Real outcomes over polished promises. Here is a flagship build, and the kind of work we take on.

Flagship build

Nemesis

Our open-source, fine-tuned offensive-security model, concrete proof of the AI engineering and security work we offer.

AI Engineering · Cybersecurity

Nemesis, a fine-tuned offensive-security model

A 27B-parameter model fine-tuned from scratch for authorised red-team work, published open-source, with zero loss of general capability.

Challenge

Off-the-shelf models refuse legitimate security tasks, blocking real red-team, CTF and research work that is fully authorised.

Approach

A 4-bit QLoRA supervised fine-tune of Qwen3.6-27B on a refusal-filtered, 7,000-row security dataset, then merged and shipped as GGUF for Ollama and LM Studio.

15/15
authorised red-team tasks passed
3/3
tool-calling tasks passed
0
capability lost vs. base model
QLoRAQwen3.6-27BGGUFMITRE ATT&CKAgentic
View Nemesis on Hugging Face
Full-cycle
Idea → MVP → support beyond
UK-based
Working with clients globally
1 business day
Typical response time
Open-source
Shipped model (Nemesis, Apache-2.0)

What we take on

The kind of work we do

Client engagements are kept confidential, so here is the shape of the work rather than names on a logo wall.

Full-stack product builds

From greenfield MVPs to production SaaS, Next.js, TypeScript, containerised and deployed on hardened, self-managed infrastructure.

Security engagements

Penetration tests, infrastructure and source-code reviews against OWASP and NIST, with clear, prioritised remediation you can action.

AI systems

Fine-tuned and adapted models, RAG knowledge bases and agentic workflows wired into real products, not demos.

Self-hosted infrastructure

This very site and its backend run on our own Traefik + Docker stack, with monitoring, rate-limiting and intrusion prevention.