// portfolio
All work
LLM platforms, specialised models and controlled agents.
What I’m building
- LLM platforms
LLM platform and agent execution
A reference architecture connecting models, tools and controls. A local lab for testing failures, recovery and action reconstruction.
Architecture and lab under development - Training · SLM
Training a small model for a SQL agent
Specialising a 4B model with LoRA, enforcing query policy and evaluating unseen tasks. Measuring gains and regressions before selecting a model.
Experimental study · three training cycles - Agent controls
ToolGuard and Qadar
Control tool calls before execution, then reconstruct the agent’s actions. Deterministic policy and analysis without exposing sensitive values.
Security previews - LLM processing
Ondine — LLM processing at scale
Process DataFrames with LLMs, save progress and distribute verification work across multiple instances.
Open source · 2,000 concurrent checks across a multi-instance fleet