DELHI, INDIA
Applied AI and security lab.
We study how models find and fix vulnerabilities, build the harnesses and tools that agents run on, and publish what we measure. We make Troy, a browser for agents, and open source tools across the stack.
What we do04
Products
Four products have their own homes: Troy, the browser for agents; Dexter, a proprietary terminal coding agent that fans work out to parallel subagents; Numera, accounting work drafted by an agent and approved by a person; and Whiteboard, a macOS workspace for command-line agents.
Open source
6 tools for agents, documents, code search and the web, among them curl_reap, terbium, trove, querion and glep. Each one keeps its source on GitHub and its documentation here.
Research
Teardowns of agent systems drawn as architecture diagrams, measured benchmark runs, and a source-by-source review of what an open-weight model can do in security work.
Bulletin
Reviews of the latest AI models and a scoreboard of independent benchmarks, with every number linked to its source.
How we work04
- Evidence attached. We exist to say what AI can and cannot do in real work. A claim nobody can check does not help with that, so each one ships with the run, the number and the source behind it.
- Built in the open. Source, documentation and release notes are public, so anyone can test our tools against what we say about them.
- The harness matters more than the model. Most of the gain in agentic coding comes from the tools, context and checks around a model. So does most of the risk, because the harness decides what an agent can reach. That is where we work.
- Small on purpose. A lab that ships fewer things can finish them and measure them properly.
A NOTE FROM THE FOUNDER
Why an AI lab
I have been building software for four years, most of it backend.
I started at Fyle, a Tiger Global-backed expense management company in Bengaluru that Sage acquired in 2025. From October 2022 to March 2025 I was on the integrations team, connecting Fyle to accounting systems such as Sage Intacct, Sage 300 CRE, QuickBooks and Xero. I built both sides of it: the backend that moved the data, and the screens finance admins used to map and export it. I came out fluent in journal entries (JEs), expense reports (ERs) and charts of accounts (COAs). That is where I learned what production means. Other people's money moves through the code you ship, and a wrong row is a real problem for a real finance team.
Velofy began in April 2025 as a small studio. I took on backend work for firms in Turkey and India: fixing slow queries, writing tests and shipping features alongside their teams. Client work teaches you speed, but it does not compound. The tools I kept rebuilding for each client did.
So the studio became a lab. Through 2026 I have also been the backend engineer at FableRoom, where I built Maestro, its operator agent, and Querion, a read-only data analyst I open sourced. Now I also work on infrastructure at Chert, a Y Combinator company that gives AI agents their own iMessage and FaceTime lines. The pattern repeated: build the thing once for real use, then publish it with the evidence.
I read a lot of what investors publish, Y Combinator (YC) and Andreessen Horowitz (a16z) most of all. Their essays taught me to ask what has changed and what that change makes possible. What changed is that models can now read, plan and act. What has not caught up is everything around them: the harnesses, the tool calls, the context they are given and the checks on what they do. That gap is where the work is.
A model that can act can also be steered, and a model that can read code can look for holes in it. Both are security questions, so security is part of the work from the start, not an extra.
That is why Velofy is a lab and not an agency or a single product. A lab can study how agent systems are built and attacked, ship the tools that close the gaps, and review the models honestly, all in public. If the work is good, it will speak for itself.
Anish Kr SinghCo-Founder, Velofy · anishfyi.com
History05
- 2022
Fyle, Bengaluru
Integrations team: backend and admin screens connecting an expense platform to Sage Intacct, QuickBooks and Xero. Fyle was acquired by Sage in 2025.
- 2025
Velofy begins
A small studio doing backend work for firms in Turkey and India.
- 2026
FableRoom
Backend and applied AI. Maestro, an operator agent, and Querion, a read-only data analyst that is open sourced, are built there.
- 2026
The lab takes shape
Open source tools with documentation, products with their own homes, the Bulletin and the research teardowns.
- 2026
Applied AI and security
In October the lab narrows its focus to applied AI and security. Projects outside that remit leave the catalog.
The people
Anish Kr Singh
Co-Founder · anishfyi.com
Based in Delhi, India. Write to anish@velofy.co, or pick a time that suits you.