Projects

Cross-repo infrastructure memory for coding agents. It knows who else depends on an artifact before you change it, and watches those pins for CVEs.

Problem

Agents are excellent inside one repo and blind across many: nothing tells you which other repos pin the base image you just bumped.

Built

Claude Code hooks that inject cross-repo impact unprompted, an MCP server on PyPI and the Claude Code plugin marketplace, and OSV monitoring filtered to the versions your pins actually resolve to.

Result

Planted 6-repo corpus: 39/39 recall, 39/39 version specs intact, 0 false positives376 tests

Install

pip install blastradius-mcp

Always-on guardian for AI coding agents. It checkpoints every file a destructive command would touch, then blocks the command, and catches secrets and agent debris before they land in your repo.

Problem

Long-running agents destroy work, leak keys, commit scratchpads, and pollute context while you’re away.

Built

A real-time filesystem daemon with a content-addressed SHA-256 journal and two-phase preflight restore, quarantine, git hooks, 22 MCP tools plus an MCP proxy, and hardened / sandbox / CI enforcement tiers.

Result

48-run ablation: block rate doubled, 0.33 → 0.67; work preserved 0.92 → 1.00; 0 false positives13 formally specified guarantees · 679 tests

Install

pip install deadpush

Open-source ChatGPT exporter that works where Settings → Export does not, including Business & Teams accounts.

Problem

Business/Teams users can’t use the official export, and third-party tools drop branches and images.

Built

A Chrome extension and a Node CLI that emit the official conversations.json, full conversation trees, and attachments.

Install

From source; not on npm.

$ git clone https://github.com/harris-ahmad/chatliberate

$ cd chatliberate

$ npm install && npm run build:extension

# then load apps/extension/ unpacked in Chrome, or run the CLI:

$ node apps/cli/bin/chatliberate.js -o ./my-backup

Go CLI for multi-repo Git workflows: clone, sync, status, branch, and local AI helpers without context-switching.

Problem

Managing many clones means repetitive pull / status / commit across directories.

Built

Parallel clone and sync, workspace-wide status / diff / stash, plus optional Ollama-powered commit, standup and ask.

Install

brew tap harris-ahmad/tap && brew install gitpull

The artifact supporting our paper at The ACM Web Conference 2024 (WWW ’24) on the data costs of YouTube ads.

Built

A public dataset and toolchain, with per-video buffer telemetry.

Data

17.6K videos · 46.6K ads · 8 countries

HIPAA-aligned e-health platform, built as a Software Engineer at the LUMS Interactive Media Lab.

Built

Patient workflows, a serverless recording pipeline, and Whisper + GPT-4 transcription into structured clinical notes (95% accuracy, 1K+ recordings/day).

Result

API latency 5s → 500ms for 20K+ daily active users