SearchBook a call

Velofy LabDelhi, India

AI research and open source tools for real work.

An applied AI lab. AI research, applied AI, and the harnesses and tool calls that make agentic coding better. We build Troy, the browser for agents, and open source tools we measure.

Book a callanish@velofy.co

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

01Products

Troy, the browser for agents.

Get Troy

A Chromium browser for macOS and Windows that you use like any other, with one difference: open the debugging port and an AI agent attaches over CDP to the tabs you are already signed into, instead of a fresh browser with no logins. It ships with a Claude Code plugin and skills that teach an agent to work in your session. Free and MIT licensed. Read the documentation.

Available · macOS, WindowsTroyYour agent joins the browser you are signed into.Early access · HarnessKestrelOne coding task, several agents, one verdict.Early access · AccountingNumeraAccounting work drafted by an agent, approved by you.Preview · macOSWhiteboardA native macOS workspace for CLI agents.Available · Open sourceSummit.jsCompose reactive behavior directly in HTML, CSP-safe.

02Signal

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.

03Papers

Research, with the evidence attached.

All 46 papers
Featured · proposalInfraze: agentic DevOps on an engine you chooseA proposal for a vendor-neutral infrastructure agent that routes each step to a cheap or an expensive engine, remembers what it built, and gates what it breaks.Read the paper →Featured · reportGLM-5.3 and open-weight cyber capability: what the evidence supportsA source-by-source reading of Anthropic's GLM-5.3 cyber report: what three parties agree on, where the framing outruns the data, and what open weights change.Read the paper →Featured · teardownOpenClaw: voice, FaceTime and where to host itHow OpenClaw does phone calls, device Talk, iMessage voice notes and experimental FaceTime, and every documented way to host the Gateway and its Macs.Read the paper →
  1. MeasuredGLM-5.3 and open-weight cyber capability: what the evidence supports
  2. MeasuredDecision Bench v0, first run: three systems on Banking77
  3. MeasuredDecision Bench experiments: temperature refit and the option-count sweep
  4. MeasuredDecision Bench experiments: near-miss distractors, per-bucket temperature, and shortlists
  1. teardown · publishedLaya: An Open Typed-Decision Engine, Examined
  2. survey · draftJev and the Typed-Decision Landscape
  3. survey · draftTyped Decisions Without Generation: A Survey
  4. survey · draftAbstention and Honest Uncertainty for Small Decision Models
  5. survey · draftWhich Encoder Backbone Should a Small Decision Model Start From?
  6. report · proposalMaestro Flash: A Non-Autoregressive Multilingual Model for Typed Decisions and Structured Outputs
Open the research graph

04Tools

Open source you can run today.

All projects
Agents · JavaScripttroyA browser built for agents to read and operate.Infrastructure · Pythoncurl_reapBrowser-grade TLS, resilient selectors, and concurrent crawling.Applied AI · PythonterbiumExtract structured data from documents, with confidence checks.Interfaces · TypeScriptsummit.jsCompose reactive behavior directly in HTML.Agents · RusttrovePersistent, file-based context for Claude Code.Applied AI · PythonquerionAsk questions of your data through a read-only analyst.

05Index

Areas

Overview
  1. Agent systemsContext, tools, and feedback for software that takes action.
  2. Applied AIUseful AI at the boundary of unstructured information and working systems.
  3. AI infrastructureThe retrieval and developer tools underneath agent workflows.
  4. Human interfacesHow people and agents interact with software, and what makes that interaction clear.

Journal

All notes
  1. 4 Oct 2026The open loop: fast decision models in cloud operationsSmall decision models are ten times cheaper and ten times faster. Whether that makes monitoring better or worse depends on a question nobody has measured yet.
  2. 3 Oct 2026The same weights are not the same modelAn endpoint is weights plus numerics, decoding, parsers, limits and operations. Each layer can change the answer. How to reason about it, measure it and buy it.
  3. 12 Sept 2026Building Velofy in the openVelofy brings together applied AI, agent systems, and the tools around them.