Ali Zulfiqar

I am a founding engineer on MillOS at Zikra Infotech, building agent-native estimation for commercial millwork. I co-founded Claimlyt, an AI platform for appealing insurance claim denials, and I have open-source fixes merged in Google DeepMind's optax, SGLang, and Rocket.Chat. I am a CS undergraduate at FAST-NUCES Lahore (Class of 2028), and I write about building software to an audience of 88k+ as codewithfourtix.

Email  /  Resume  /  Github  /  LinkedIn  /  Devpost  /  Instagram

Things I have built, shipped, or broken enough times to understand.

An agent-native platform (LangGraph, a lead estimator plus 8 subagents) that turns architectural drawings into commercial bids. I rebuilt the measurement engine to 0.42 percent live accuracy against a 2 percent bar, built the annotation engine that places marks with median error of 0.1 inches, and own the estimation studio frontend and much of the agent reliability layer.

A platform that automates appeals against insurance claim denials for US healthcare organizations: a seven-node LangGraph pipeline with pgvector retrieval that grounds every appeal in similar successful cases. Applied to Y Combinator (S26). Co-founder.

An Urdu voice agent for blind and low-literacy Pakistanis that checks money, medicine, and documents. Its camera loop triages frames in the browser, verifies medicines against a saved prescription, and speaks warnings before a user commits to a document. Built for the Uplift, YC-backed company × Replit Hackathon.

A browser game where you roam actual streets, draw loops to claim ground, and steal rivals' turf. Server-authoritative multiplayer, framework-agnostic canvas engine over Leaflet, Playwright smoke suite. Code.

Runs Qwen2.5 on CPU with a hand-written transformer forward pass (RMSNorm, RoPE, grouped-query attention with KV cache, SwiGLU) and custom INT8/INT4 quantization, 4x and 7x smaller with INT8 quality-lossless by perplexity. No ML framework; verified against a NumPy reference oracle.

A full autonomous-vehicle stack in the browser: Frenet lattice planning, Kalman multi-object tracking, Stanley control, and neural drivers trained from scratch with no ML library (backprop plus Adam, imitation, DAgger, neuroevolution). Deterministic tests assert zero collisions across nine scenarios. Code.

A transformer implemented from first principles in plain JavaScript, written to teach LLM architecture with zero dependencies between you and the math.

A Go MCP server that connects Claude to any project directory, serving structured code context over the Model Context Protocol.

A Claude Code skill that strips the tells of machine writing: filler phrases, em-dash chains, hedging, and symmetric slop.

200+ users in the first month via a Product Hunt launch. A hybrid OCR-first pipeline cut LLM token cost 70 percent and multimodal batching cut API overhead 96 percent, at sub-2s processing. Devpost.

An always-on-top, edge-docked todo strip for Windows. WPF, zero dependencies, 25 kB.

Caught a numerical-correctness bug in Google DeepMind's optax, confirmed by a maintainer as a genuine problem with the implementation; fixed an LLM tool-call parser in SGLang; landed fixes in Rocket.Chat and scancode-toolkit. 10+ PRs across open source.

Real performance metrics for semantic search embedding models, measured rather than quoted.

Three awards, a top-25 finish, and an audience built in public. Devpost.

An AI credit-origination layer for Pakistani SMEs: applicants apply through WhatsApp in Urdu, while bank officers receive an explainable pipeline that extracts documents, verifies inconsistencies, scores repayment risk, and produces an auditable one-page brief.

A verification tool for WhatsApp users that identifies scam messages and voice notes and returns verdicts in Urdu, built for the country's largest civic hackathon.

Your AI COO on a link anyone can use: an agent that runs front-of-house operations end to end.

The AI document processing tool that later shipped as getzippo.com with 200+ users in its first month.

I write about building software, AI systems, and shipping projects in public.

A growing professional audience around software engineering, startups, and building in public.

Published work.

A vision-language model names the items and reads the printed dimensions; a deterministic geometry engine recovers the architect's own rectangles from the vector layer and places every mark by arithmetic. Six of six casework marks exact on a production sheet, median width error 0.1 inches.

Shifts System-2 reasoning to asynchronous idle cycles, making response latency independent of reasoning depth.

Page design adapted from Jon Barron's website.