All work
AI AgentsResearch2025
Agent red-teaming framework
A dual-agent system that attacks safety-aligned models on purpose, then measures exactly how and when they break.
Scope something similarCode walkthrough available on request
The problem
Before you put an agent in front of customers, you need to know what it does under adversarial pressure, not hope for the best.
What we built
- Persuader and Persuadee agents driven by LangGraph, with stateful multi-turn conversations and conditional routing
- Fourteen jailbreak techniques across four strategy families, run in both single-turn and multi-turn modes
- Automated safety judge (Llama-Guard) classifying every response, with human validation on samples
- Experiment store in PostgreSQL via SQLAlchemy with Alembic migrations, exposed through a FastAPI service
- Streamlit and Plotly dashboards reporting success rates and normalised change per strategy and model
Need something like this?
We'll tell you honestly whether your case is a two-week build or a three-month one — and what it would cost — before you commit to anything.
More in AI Agents
InfoTech Wizard — AI IT-support desk
A RAG chatbot that answers real IT-support questions from a company's own documentation instead of guessing.
Production AI agent workflows
Four business agents where the LLM does the reading and deterministic code does the checking: invoices, leads, content and support.
