I build systems that test other systems.
Independent engineer working on open-source models, local inference, and security tooling. Currently building an open cybersecurity agent and the harness it runs in.
Writing
All posts →- What I am building: an open cybersecurity agent
A model and a harness for security work, released open source. Written down early so the claims are checkable later.
- The decade of the agent
Agents are oversold on timeline, not on direction. The gap between a demo and a system that survives a bad Tuesday is where the whole argument lives.
- The bottleneck moved
AI did not kill software engineering. It made writing code cheap, which relocated the hard part to deciding what to build and proving it works.
- Cheaper thinking, more work
The mass-unemployment argument assumes demand for cognitive work is fixed. Every previous time we made something dramatically cheaper, it was not.
Work
Work →- Models · Prisma
Prisma
A small open-source organisation I run. It trains and releases open models. Everything is built and maintained by one person, which sets the constraint: small models, tight data, no budget for waste.
- Inference
Local LLM stack
A local inference setup built on llama.cpp with my own model and provider configuration, plus Qwythos on top. Running models on my own hardware taught me more about their real limits than any benchmark did.
- Security
Red-team environment
A test environment for attacking my own software and local endpoints before anyone else does. Fuzzing, prompt injection, misconfigured services, exposed local ports. Everything is run against systems I own.
Now
Building a cybersecurity agent model and its harness. Both will be released open source under Prisma, a small open-source organisation I run that trains open models.