SearchBook a call

Applied AI and security research for real work.

Velofy is an applied AI and security lab in Delhi, India. We study how models find and fix vulnerabilities, build the harnesses and tools that agents run on, and publish what we measure.

Papers
46
Measured runs
04
Open questions
70
Open source tools
06

Products05

Get Troy

TroyAvailablemacOS, Windows

Your agent joins the browser you are signed into. A Chromium browser for macOS and Windows: open the debugging port and an AI agent attaches over the Chrome DevTools Protocol (CDP) to the tabs you already use. Ships with a Claude Code plugin and skills. Free and MIT licensed.

DexterProprietaryTerminal

A terminal coding agent and multi-agent harness. When a job splits into independent parts, the model fans them out to parallel subagents on its own, up to 8 at once, each in a live tmux pane. Runs on any model, from GLM-5.3 and Muse Wonder Full to local weights, learns from every fix, and keeps working under a spend cap when you step away.

NumeraEarly accessCommercial real estate

One close for every property: agents draft the month-end across every property and project entity, you approve each entry, and Numera reads it back from the ledger and seals it.

WhiteboardPreviewCloud

The shared workspace for people and agents: channels organize the conversation, tickets carry ownership, and every change reaches you as a real diff to review.

InfrazeEarly accessDevOps

State the intent and Infraze runs the stack: Kubernetes, Docker and Grafana in parallel, with anything destructive stopped for your approval.

Dexter · Proprietary

Parallel subagents, up to eight at once.

Dexter is Velofy's terminal coding agent and multi-agent harness. Point it at any repository and any model, and it plans, edits, runs and verifies. When a job splits into independent parts, it hands them to parallel subagents on its own and shows each one working in a live tmux pane.

Visit dexter.velofy.co
  1. Parallel by defaultThe model splits independent work into subagents without being asked: up to 8 at once, each in its own process group. Esc stops them all.01
  2. Live panes anywhereEvery subagent streams into its own tmux pane, inside tmux or from a plain terminal tab.02
  3. dexter fanoutScripts, continuous integration and other agents run the same swarm from a JSON task list, and a bundled Claude Code skill hands work to it.03
  4. Any modelGLM-5.3, Muse Wonder Full, Gemini Argon, Anthropic, OpenAI, or local weights through Ollama and llama.cpp.04
  5. Learns and healsMemory carries across sessions, long sessions become reusable skills, and every fix proven by a passing test becomes a playbook.05
  6. Keeps workingIts keepalive loop walks your ROADMAP.md under a rolling spend cap and a one-file kill switch, and the vortex loop evaluates, fixes and re-runs until a goal is met or its budget runs out.06
  7. 200+ skills and a research kitiOS, macOS, Android, GitHub, design, testing and debugging, plus built-in repository search, document parsing, browser-grade fetching and browser control.07
  8. Lean at startupOne copy of dexter per process: 1,652 files loaded at startup instead of 2,853.08

The Velofy ecosystem

Built on Dexter. Runs natively in the cloud.

Every Velofy product runs on Dexter, the lab's multi-agent harness: Infraze for DevOps, Whiteboard for shared work with agents, Numera for commercial real estate, and Troy, the browser agents join. One harness underneath, so each product gets parallel subagents, any model and the same guardrails, and all of it runs natively in the cloud.

  1. DexterThe multi-agent harness underneath every product: parallel subagents, any model, self-healing playbooks.01
  2. InfrazeThe DevOps agent: Kubernetes, Docker, Grafana and the rest of the stack, run from a stated intent.02
  3. WhiteboardThe shared workspace where people and agents plan, run and review work.03
  4. NumeraCommercial real estate: the month-end close across every property, proven line by line.04
  5. TroyThe browser your agents join, signed in to everything you are.05

Signal

A trained baseline sets the bar at 77 labels.

On a 77-way Banking77 slice, a TF-IDF model trained on 10,003 examples answers in 0.19 ms and beats both zero-shot decision models. At 5 options the picture changes, so option count is a first-order variable.

Read the report
Fig. 1 Accuracy, 1,001 items, 77 options. Higher is better.
  1. TF-IDF + LRtrained0.846
  2. Layazero-shot0.386
  3. GLiClass-smallzero-shot0.119

Random guessing: 0.014. Source: Decision Bench v0.

Research46

All papers

Open the research graph

Open source06

All projects

Areas04

Overview
  1. Agent systemsThe context, tools and checks around a model that takes action.01
  2. Applied AIModels that turn documents and questions into results you can check.02
  3. AI infrastructureThe code search and web retrieval tools that agent work runs on.03
  4. AI securityHow models find and fix vulnerabilities, how agent systems are attacked and defended, and what a network fingerprint gives away.04

Journal09

All notes
  1. Fingerprints below HTTP: how servers recognise the software that connectsHow servers recognise clients by TLS ClientHello, HTTP/2 settings and header order, how defenders use it, where it fails, and what it means for agents.
  2. Least privilege for AI agents: put the limits where the model cannot reachA prompt can ask an agent to behave; roles, tokens, network rules and sandboxes make it. How to bound an agent, and where a human must still confirm.
  3. Prompt injection is a confused deputy problemA model reads instructions and data on one channel, so an agent holding authority can be steered by what it reads. The designs that hold by construction.
  4. What an agent harness is made ofAn agent is a model plus a harness: loop, tools, context, checks, permissions and state. Why harnesses change scores, and how to compare them.