For researchers
And gets published. Reinforcement learning, LLM verification, domain ontologies, and the engineering behind production-grade agentic AI – the work that lets agents earn the right to operate in your environment.
Research · from the Nablon stack
Reinforcement learning · Featured paper
11× cheaper than GPT-4o · runs on-prem
A GRPO-trained Llama 3.1 8B policy matches GPT-4o at 83.4% F1 on 58 PII entity types, at 11x lower cost, running entirely on your infrastructure. The reward function preserves spans and penalizes leakage. No customer data ever leaves the environment.
Read the paper →What we work on
The model is swappable. The shell around it – ontology, rubrics, decision traces – is what we build, and what makes an agent yours.
More on research
Expert knowledge encoded as a Dirichlet prior, root causes ranked by posterior confidence, and a production feedback loop that updates belief from confirmed outcomes. No model retraining.
A meta-agent system and visual IDE that turns a description, a PDF, or a PRD into an inspectable, runnable multi-agent architecture – and exports it as a GitHub-ready Python project.
Moving beyond static policies. A production system that rewrites its own prompts in real time, validates them automatically, and compounds improvements across sessions – no human in the loop.
5× cost reduction through prompt bundling. 71 scanners across 5 execution mechanisms. A full lightweight evaluation path under 500 ms.
Partners
Investors
