AI/ML Engineer, Researcher & Systems Architect
I build AI systems that have to work in production: voice agents that answer real phone calls, RAG and LLM agents judged by cost per correct answer, and multi-tenant SaaS platforms with row-level security and field-level encryption. I also publish the research behind them, including the numbers that went against me, and open-source tools that catch problems before they ship: Microsoft gov-cloud gaps, broken Make.com scenarios, Postgres images that crash on older CPUs and packages an AI coding agent made up. Based in Pakistan, working with US teams, open to contract work.
28 public repositories, grouped by what they are for, newest work first. Descriptions come straight from each repository; the green line under a tool is something you can check.
Catch the problem before it ships.
Is it available in GCC High? Check a Microsoft 365 GCC, GCC High, DoD or Azure Government plan against Microsoft's own availability tables: 633 cited features, 57 endpoints, a wrong-cloud host scanner, MCP server and GitHub Action.
On PyPI and the GitHub Marketplace. Re-checks all 33 Microsoft sources weekly and keeps a Wayback Machine copy of each; on 12 Oct 2026 it caught Microsoft rewording two rows on the Teams page.
GitHub · Website · PyPI · GitHub Action
Test Make.com (Integromat) scenarios offline: run exported blueprints locally, replay real Make runs and catch broken mappings before you deploy. CLI test runner and blueprint validator for CI. 71 of 79 real Make runs reproduced exactly, 0 differ.
On npm. 71 of 79 recorded real Make.com runs reproduced exactly, 0 differ; the other 8 are refused instead of guessed.
Find out which CPU crashes your Postgres image before your users do. Runs it on 7 CPUs (2008 Intel to Raspberry Pi and Graviton2) and names the faulting instruction (SIGILL). GitHub Action.
On the GitHub Marketplace. 13 popular Postgres images audited on 7 CPU models, 3 build bugs found.
Attackers register the packages LLMs invent. diff-gate reads your coding agent's git diff and fails on imports that resolve to nothing. Claude Code skill + GitHub Action, zero dependencies.
On the GitHub Marketplace. Zero dependencies.
Bitnami's PostgreSQL and PostgreSQL HA (repmgr) images rebuilt from upstream source: same paths, scripts, env vars and tags, drop-in for the Bitnami postgresql and postgresql-ha Helm charts. No Broadcom binaries, every source SHA-256 pinned, amd64 + arm64.
On the GitHub Marketplace. Every upstream source SHA-256 pinned; amd64 and arm64 images.
Free MinIO Docker image for 2026: MinIO server + mc built from source at pinned, commit-verified releases. Multi-arch (amd64, arm64), smoke-tested replacement for minio/minio and quay.io/minio. Unofficial.
On the GitHub Marketplace. Multi-arch images on GHCR, built from commit-verified releases.
Measured, with the numbers that went against me left in.
Open reproduction of TypeSafe Jev: a 150M typed decision engine (noul/choice/score in one non-autoregressive pass, calibrated confidence). 0.697 vs Jev's 0.727, 2.5x better calibrated, 4x faster, free. Trains on a Colab T4 in 30 min.
46 stars. Trains on a free Colab T4 in about 30 minutes; 0.697 accuracy vs TypeSafe Jev's 0.727, with the loss reported up front.
What does agentic RAG actually cost? A measured cost-per-correct-answer benchmark of agentic RAG, iterative RAG, cross-encoder reranking and naive RAG on multi-hop QA (MuSiQue). Meters every LLM call.
Agentic RAG scored +32.5 exact match over naive RAG and cost 2.5x more per correct answer.
9 PDF-to-Markdown parsers for RAG on a plain 4-core CPU: olmocr-bench score, seconds per page, peak RAM, install size. MinerU's CPU tier leads at 74.6%.
Quality scored on 1,403 PDFs with the official olmocr-bench checker.
Prove which AI fetched your page: secret canary links, a Cloudflare Worker log and published-IP checks. Tested on ChatGPT, Claude, Gemini, Grok, Perplexity and Copilot (Oct 2026).
Every request log, prompt and result is in the repo, timestamped with OpenTimestamps.
How often do you correct your coding agent, and on what kind of work? Measures it from your Claude Code transcripts, drafts CLAUDE.md rules from your recurring corrections, and exports them as preference pairs.
On PyPI, with tests in CI.
Crypto trading strategy backtesting & ML validation: triple-barrier labels, purged k-fold CV, deflated Sharpe ratio and a positive control. LightGBM, LSTM/BiLSTM, Binance + MEXC perpetual futures data.
Voice agents, RAG assistants, lead pipelines and ML APIs.
CRM & call-ops backend for Vapi AI voice agents: webhook ingestion, predictive dialer, auto-redial, payment-verified conversions (Stripe + portal reconciliation), Twilio call status and analytics. Node.js, Express, SQLite/MySQL.
Self-hosted lead generation engine: scrape local business leads (phone, website, e-mail) past the Google Places 60-result cap with grid search, free OpenStreetMap fallback, dedup, adaptive rate limiting & resumable runs. FastAPI + Streamlit.
Healthcare lead generation from the NPI Registry: fetch providers via the NPPES API, find practice websites, extract e-mails. Resumable Python pipeline scheduled on GitHub Actions — zero paid APIs.
AI financial advisor chatbot: RAG + tool-calling LLM agent for portfolio risk (VaR, Sharpe, drawdown), diversification scoring & investment strategy, with ChromaDB memory. Gemini, OpenRouter, LangChain, Flask, Microsoft Fabric.
Face recognition access control with liveness detection & anti-spoofing: InsightFace ArcFace embeddings, head-pose challenge, rPPG pulse detection and ChromaDB vector matching in a Streamlit dashboard.
Crypto price prediction API with a PyTorch BiLSTM + attention model: 30m/1h/4h/24h forecasts for BTC, ETH & altcoins, CoinGecko + Yahoo Finance data, WebSocket streaming and JWT auth. FastAPI.
Small things that save someone an afternoon.
libaio1, libncurses5, libncurses5-dev & libtinfo5 .deb packages (amd64) — fix 'libncurses.so.5 / libtinfo.so.5 / libaio.so.1: cannot open shared object file' on newer Ubuntu.
Working notes on git internals and GitHub Actions
Learning projects from before the work above, kept public and archived.
v1 of CryptoAion: FastAPI service forecasting Bitcoin, Ethereum & altcoin prices with a PyTorch BiLSTM + attention model, CoinGecko and Yahoo Finance data, plus BiLSTM and XGBoost research notebooks.
Flask microservices on Kubernetes (Minikube + Flannel) and Docker Swarm: NGINX reverse proxy, PostgreSQL, logging service, Prometheus monitoring and NetworkPolicies — a hands-on DevOps project.
Flask + PostgreSQL CRUD app deployed on Kubernetes (Minikube) behind an NGINX reverse proxy with NetworkPolicies, plus a Docker Swarm stack.
Flask + PostgreSQL user management CRUD app with a Bootstrap dark UI, Chart.js dashboard and REST API — runs with Docker Compose.
Full-stack TypeScript CRUD app: React + Vite, Express, Drizzle ORM, Neon Postgres, Passport session auth with scrypt hashing and a Postgres session store.
DevOps & cloud services agency website: React + TypeScript landing page with Tailwind & shadcn/ui — services, pricing calculator, testimonials, blog, FAQ, live chat — and an Express contact/newsletter API.
Interactive Bash script to install or update Python, VS Code, Docker, Jenkins and kubectl on Ubuntu — detects installed versions and asks before changing anything.
Array-based FIFO queue in C++ with enqueue, dequeue, size and display — a data structures coursework exercise.
Every project card on my GitHub profile is hand-coded, animated SVG in the style of Pakistani truck art.
I design systems end to end, from the data model to the deploy pipeline. Patterns I use in production:
I take on contract and freelance work: