AKAnimesh Kundu
Menu

ANIMESH KUNDU / PUBLIC WORK INDEX

Built in public, from utility to judgment.

A Firefox extension used by 14,484 people each day. A small command model trained from scratch and shipped inside a Rust CLI. Agent systems, browser tools, infrastructure, and a steady record of essays.

Inspect the workRead the essays

EVIDENCE / DATED

Proof before promise.

Each record leads to a public source. Changing numbers carry the date on which they were checked.

E-0114,484 daily users

youtube-audio keeps video off when only the sound matters. Mozilla also reports 527 ratings and 224 written reviews.

Source, Jul 21 2026
E-02135K training examples

cmd-correct is a Qwen3 LoRA trained with MLX and published as a compact GGUF model behind the offline fix CLI.

Model card
E-0340 published essays

Notes on agent work, organizations, attention, privacy, cities, health, and the systems around everyday choices.

Feed snapshot, Jul 22, 2026
E-0431 durable repositories

A local snapshot keeps the record present even when GitHub is unavailable or rate-limited.

Snapshot, Jul 22, 2026

SELECTED WORK

Different layers, one habit.

Find the useful edge, make the whole path work, and leave the source open for inspection.

See all project records

W-01 / Shipped products

RUNNABLE

YouTube Audio

A Firefox extension that turns off YouTube video playback to save bandwidth and battery.

  • 14,484 average daily users reported by Mozilla Add-ons
  • 527 ratings and 224 written reviews
Read the recordSource

W-02 / Models and agents

SOURCE

Fix

A Rust command correction CLI powered by a small, self-trained model that runs locally.

  • No API key or hosted inference dependency
  • Published through Homebrew, WinGet, and Cargo
Read the recordSource

W-03 / Shipped products

SOURCE

Oops

A compact Rust command-line typo corrector built for quick feedback.

  • Rust implementation
  • Fast local correction
Read the recordSource

W-04 / Models and agents

SOURCE

Condukt

A composable workflow framework for agent execution, state, orchestration, runtimes, and UI.

  • Separates workflow definition from runtime
  • Treats state and orchestration as first-class concerns
Read the recordSource

W-05 / Models and agents

SOURCE

GitHub Router

A local router for using Claude Code, Codex CLI, and related coding tools with a Copilot subscription.

  • Connects multiple coding CLIs to one subscription
  • Public TypeScript source and live Pages surface
Read the recordSource

W-06 / Systems and libraries

SOURCE

Runnerize

A dispatcher that mints an ephemeral JIT GitHub Actions runner for each queued job in a throwaway rootless container.

  • One runner per queued job
  • Rootless container isolation
Read the recordSource

MODELS + AGENTS

The system around the model.

Training is one layer. Routing, state, execution, interfaces, and verification decide whether the result becomes useful.

A-01Fix

A Rust command correction CLI powered by a small, self-trained model that runs locally.

Open record
A-02Condukt

A composable workflow framework for agent execution, state, orchestration, runtimes, and UI.

Open record
A-03GitHub Router

A local router for using Claude Code, Codex CLI, and related coding tools with a Copilot subscription.

Open record
A-04Runnerize

A dispatcher that mints an ephemeral JIT GitHub Actions runner for each queued job in a throwaway rootless container.

Open record

WRITING / Jul 22, 2026

Questions that survive the build.

The essays follow consequences past the demo: what agents cannot infer, what organizations lose, and what convenience asks in return.

Browse the writing index
N-01

Never Left Hanging

AI chatbots can offer comfort when human care is scarce, but their deeper risk is teaching us to prefer understanding without delay, pushback, or the burden of another person.

N-02

No One at the Next Desk

AI-native startups can now begin with one founder and many agents, but the missing early teammate may be the human witness who catches bad taste, bad assumptions, and dangerous commands.

N-03

The Meeting Grew a Memory

AI notetakers can make meetings fairer by giving rooms a record, but the same memory raises the cost of unfinished, candid speech.

N-04

What the Agent Can't Guess

Coding agents are making specs matter again: not as old process theater, but as the durable statement of intent that generated code must prove or betray.

BROWSER TOOLS

Open the file, not an account.

Inspect a network capture, edit a diagram, read a sequencing trace, or work with a PDF. Direct links remain available if an embed cannot load.

Try the tools
T-01Mermaid Editor

Write, preview, share, and export diagrams.

Open tool
T-02Sanger Viewer

Read chromatograms and base quality in the browser.

Open tool
T-03HAR + SAZ viewers

Inspect network captures without a desktop install.

Open tool
T-04PDF Viewer

View, annotate, and export a document locally.

Open tool

NEXT CONVERSATION

Bring the unfinished question.

Complex product, model, or systems work usually starts before the shape is obvious.

If a public artifact here connects to something you are building, send the context, the constraint, and the part that still feels uncertain.

Start with context