Velofy is an applied artificial intelligence (AI) and security lab in Delhi, India. This site collects the lab’s public work: tools, measured research and notes.
First principles
An agent is a model plus everything around it: the harness that runs it, the tools it calls and the data it reads. What the agent does depends on all of these parts, so studying the model alone misses most of the system.
Each of those parts can be attacked. The model can be steered by what it reads, and the harness and tools are software acting on untrusted input. An agent that reads the web or runs code takes in untrusted input all the time. Its security is therefore a property of the whole stack, and it belongs in the same lab as the agent work.
A claim about a model or a tool is only as strong as the evidence behind it. Evidence that other people can rerun is stronger than evidence they have to take on trust. That is why the lab publishes its source code, its methods and its measurements.
Together these set the work: build the tools agents run on, test how they hold up and how they break, and publish both.
The work
Troy is a browser for agents. Trove keeps file-based context between sessions. Querion is a read-only analyst for connected data. Terbium extracts structured data from documents, and curl_reap reads the web.
The AI security research area covers how models find and fix vulnerabilities and how agent systems are attacked and defended.
What comes next
When this post went up, the plan was for products to follow the open source work, with details once they were real.
Update, October 2026: the products now have their own sites. Troy is the browser for agents. Dexter is a proprietary terminal coding agent that fans work out to parallel subagents. Numera drafts accounting work for a person to approve. Whiteboard is a macOS workspace for command-line agents.
The open source index is the place to start, and the research papers carry the evidence behind the work.
Questions? Email Anish or book a call.