Abdul Rafay Zahid

I scale C++ network gateways, build high-volume mail infrastructure, and write the systems plumbing under AI products. About three years in, and happiest when the problem is below the framework layer: memory, connections, queues, and the failure paths that only show up under load.

Three things I have shipped

A C++ TLS WebSocket gateway taken from 200 to 5,000 concurrent clients. Replaced blocking polling, pooled connections, batched Arena requests, and initialized Arena memory up front. Then stress-tested it past normal traffic and fixed what only broke there, including a log-rotation race that crashed a worker thread.

Enterprise email infrastructure sending 200K+ emails a day across 1,000+ domains, built solo. A campaign engine on KumoMTA and Amazon SES edge nodes, automated SPF, DKIM, and DMARC lifecycle, Redis-locked warmup, blacklist monitoring, bi-directional Salesforce sync. It beat the Salesforce tool it replaced on response rates.

Colloid Swarm, a portable multi-agent engineering harness now used company-wide by roughly 50 engineers. Seven delegating personas across three cost tiers, hostile review as an independent gate, one engine-neutral layer with thin adapters for Claude Code, Codex, Kimi, and Copilot.

Where I have worked

Currently a Software Engineer at Softaims. The work has ranged from the email platform above to a voice onboarding agent on ElevenLabs Conversational AI whose backend keeps conversation state in a slot-walk, so the whole interview script replays offline without a live call, with scoped tool tokens, signed webhooks, and exactly-once writes. Also a LangChain deep-research agent framework, a RAG recommendation system on OpenAI embeddings, and Playwright scraping infrastructure collecting 10,000+ records a day from 50+ sources.

Before that, at Ebryx, I worked on a zero-trust network access platform in C++ and Python: the gateway scale-up, an auto-update migration system for a multi-tenancy move, a ZTNA component ported from x86-64 to ARM for on-prem deployment, MySQL pooling rewritten with RAII, policy evaluation re-architected from cron to event-driven (29% less redundant database load), and a Kafka pipeline scoring network-agent risk for both ML training and live inference.

Earlier, OCR and Elasticsearch work at Active Takeoff, and internships at Systems Limited and PITB. The full timeline is on the Experience page.

What I work with

Daily: Python, C++, TypeScript, SQL, Bash. FastAPI, NestJS, Next.js. PostgreSQL, MySQL, Redis, Kafka. Docker, Linux, Terraform, AWS, GCP.

Comfortable: LangChain and LangGraph, MCP, OpenCV, NumPy and Pandas, Cython, Playwright, Elasticsearch, MongoDB.

Learning: Rust, by building a terminal music player.

Education

BS Computer Science, FAST NUCES, Lahore, 2020 to 2024. Coursework in artificial intelligence, computer vision, parallel and distributed computing, and advanced database concepts. For my final-year project I built a real-time computer vision pipeline in OpenCV that scores how well a user mimics a facial expression, on top of an emotion-recognition CNN trained from scratch on the 28,709-image FER2013 training set and exported to ONNX for on-device inference inside a Unity game. The write-up is here.

Outside the terminal

Deputy Head of Conferences at Softec, FAST NUCES, in 2022: hosted the largest tech conference in Pakistan and ran the volunteer team for it. Before that, founded the World Affairs Society at Karachi Grammar School and launched its flagship event, Renaissance 19.