I'm Haroon, an engineer focused on turning difficult ideas into reliable products.

Available for work

Private AI, local-first tools, and reliable backend systems — Android to Go to Python, with the work shared in open source.

Open-source contributions

Work visible in public.

Pull requests

Merged

Repositories

Organizations

Signal / field

Move the pointer

Selected work

Built to work in the real world.

Wake AI

Private, searchable AI memory built for Android.

Kotlin, Jetpack Compose, Gemma

View code

Problem: Useful information is scattered across notifications, screens, apps, and conversations, then becomes difficult to retrieve when only its meaning is remembered.
System: Wake captures text on device, stores it in Room, combines FTS keyword retrieval with local embeddings, and grounds answers in cited memories using an on-device or user-selected cloud model.
Engineering proof: On-device text and embeddings
Hybrid semantic and keyword retrieval
Cited answers with approval-gated actions

How it works

Capture on device
Room · FTS · embeddings
Hybrid retrieval
Cited answers

DeadDrop

A local-first coding task inbox reachable from anywhere.

Go, FastAPI, PostgreSQL

Live demo View code

Problem: Remote task handoff should not require exposing a developer machine, local repositories, or credentials to a hosted service.
System: A hosted FastAPI inbox queues work while an authenticated Go worker polls outbound, runs the configured local agent, and returns logs, status, summary, and git diff.
Engineering proof: Outbound-only local worker
Token-separated owner and worker access
Human-reviewed diff receipts

How it works

Browser inbox
FastAPI queue
Go worker (outbound)
Local agent → diff

OutSync

Reliable database event delivery through the transactional outbox pattern.

FastAPI, Kafka, PostgreSQL

View code

Problem: Writing business data and publishing its event separately creates a failure window where state changes can reach the database but never reach the queue.
System: The application writes business data and an outbox record in one transaction; an asynchronous worker then publishes pending events to Kafka and marks them processed.
Engineering proof: Transactional outbox
Asynchronous polling worker
Containerized Kafka and PostgreSQL stack

How it works

App transaction
Outbox record
Polling worker
Kafka → processed

Herald

A local-first communication agent that sends only after explicit approval.

Rust, SQLite, SMTP / IMAP

View code

Problem: A communication agent with your account is a program that can speak as you. If the model can touch the provider directly, every send is a trust violation.
System: Herald keeps authority in Rust: the model emits a typed ActionPlan and never touches a provider, a rules router and small local classifier do the planning, and every send becomes a pending approval with body and SHA-256 that only `herald approve` can send.
Engineering proof: Approval with body + SHA-256 before every send
Idempotent writes with audit log
Local model planning, no cloud inference

How it works

Read providers locally
Draft with local model
Pending approval (hash)
Approve → send

Zyquor

A local-first X posting assistant with grounded drafts and a review queue.

Rust, TypeScript, SQLite

View code

Problem: Posting well is the hard part. Most AI social tools write generic filler, decide what to post for you, and automate engagement.
System: Zyquor reverses that: the LLM writes candidate text only, deterministic quality rules reject slop, every draft lands in a review queue, and a worker posts approved jobs through a dedicated Chrome profile.
Engineering proof: LLM writes candidates only, never browses or schedules
Deterministic quality rules over LLM opinion
Review queue gates every publish

How it works

Idea / source item
Grounded candidates
Review queue → approve
Worker → Chrome profile → post

Everything else

Containerized Agent

Natural-language shell and Python task execution in isolated Docker environments.

Docker, FastAPI, Python

CrawlWise

High-performance crawling and structured extraction across thousands of domains.

TypeScript, React, AI/ML

Code

PromoAgent

Autonomous community outreach with tailored, high-engagement content.

Python, LangGraph, Marketing

Code

Percolation Hypotheses

A hypothesis generation engine built to accelerate computational research.

Python, NLP, Research

Code

Neural Network From Scratch

Deep learning fundamentals implemented from first principles.

Python, ML Fundamentals

neighbourd

LAN discovery and pairing for local AI agents. No broker, no cloud.

Go, mDNS, MCP

Meeting Scheduling Agent

Local scheduling agent with LLM intent extraction and deterministic policy, booking real Google Calendar events.

Python, Google Calendar, OpenRouter

Notion Agent Bridge

Chrome sidepanel to localhost bridge to agent-CLI edits applied to the Notion page DOM.

Python, Chrome Extension, opencode

Those times when you don’t feel like working, you’re too tired, you don’t want to push yourself, but you do it anyway.That is actually the dream. That’s the dream.

Kobe Bryant, 2017

Read résumé Open-source contributions

Fig. 01Terrain, drawn one pixel at a time in your browser.