Projects

AgentForge

A Gauntlet AI project

A clinical co-pilot inside OpenEMR that answers a covering physician's questions about the patient whose chart is open.

Built into the open-source OpenEMR health record as a module, with the agent in its own TypeScript service. Every clinical claim must cite a record in that patient's chart, and rules enforced in code plus a second model check remove any claim they can't verify instead of hedging it. The agent can only read the one chart already open, through a short-lived scoped token, and its traces stay in self-hosted Langfuse.

AgentForge Adversary

A Gauntlet AI project

A red team for the clinical co-pilot that attacks it every night and on demand.

A multi-agent adversarial evaluation platform for the OpenEMR Clinical Co-Pilot. It generates and mutates attacks, has an independent judge rule on each one, files confirmed exploits as tickets, writes vulnerability reports and re-tests fixes as regressions, all inside a nightly budget. Its test suite is mapped to OWASP categories, and operators queue work and track tickets from a dashboard.

AI Agent Orchestration Platform

A Gauntlet AI project

Build multi-agent workflows in a web UI, with a human approval before any side effect.

Configure AI agents (prompt, model, tools, limits) and wire them into multi-agent workflows in a visual builder, with no code changes. Agents run on the open-source Goose runtime, but every control lives in the platform: tool allowlists, spend and step limits, cancellation, and an approval gate that runs an approved action exactly once. Every run is traced and streamed live, quality is measured by an offline evaluation harness, and an agent can be chatted with from Telegram.

AI Onboarding Assistant

A Gauntlet AI project

Guided sign-up that swaps long forms for a conversation and retyping for a photo of your ID.

A chat assistant collects what the service needs and saves it as a structured assessment, a photo of an ID pre-fills identity fields for the user to confirm, and a scheduler books the first appointment. ID images are read by self-hosted OCR and never sent to an LLM, and consent, data minimisation and verifiable deletion are built in. The LLM provider is swappable and was chosen by running the same evaluation on every candidate.

AI Call Center Agent

A customer-service chat agent that answers from the company's own documents, books appointments, and hands everything else to a person.

Built for a residential heating, cooling and plumbing company. Answers about plans, warranties and policies must cite the documents they came from, or the agent hands the conversation to a human with a note they can act on. It books, moves and cancels appointments in the scheduling database, looks up balances, and screens every message for safety issues like a gas smell before the model sees it, while an evaluation harness replays held-out real calls to measure it.