I'm studying cyber security at Air University in Karachi. Most of what I know about building software comes from hackathons and internships. In 2026 I built voice and LLM apps for competitions run by USAII, lablab.ai, OpenAI and UNESCO, spent the summer as a machine learning intern at FlyRank AI, and sent bug fixes to litellm and elizaOS. I'm now a community manager intern at lablab.ai.
Security is the part of AI engineering I want to work in. When a model's output decides something, I add a check in code that the model can't override. In BISP Benefits Navigator a rules engine makes the eligibility call, and in codexfoundry a code change from Codex needs a person's approval and a separate verifier before it counts. I'm on the Blue Team at the AU Cybersecurity Society.
Built solo OpenAI Build Week, July 2026
A local tool that runs Codex, OpenAI's coding agent, inside a gated loop. A planner first writes a plan, and nothing runs until a person approves it. The code change then happens in a disposable git worktree pinned to a known commit, and a separate verifier has to confirm the fix before the result is written back to the project's notes.
The recorded reference run fixes a checkout bug that charged customers twice. The repo has 79 tests.
Top 10 finalist, built solo USAII Global AI Hackathon 2026
A voice-based pre-screening tool for Pakistan's Benazir Income Support Programme. Applicants describe their household in Urdu and hear an eligibility estimate read back to them.
Gemini only pulls income, family size and assets out of the conversation. A separate TypeScript rules engine turns them into an estimated Proxy Means Test score, so the model can't change the outcome. The score is a mock estimate for the prototype, not the government's formula. The project ranked 9th of 424 teams in the qualifiers with 96 out of 100, and scored 66.5 out of 70 in the final.
“BISP Benefits Navigator removes literacy, language, and access barriers that prevent vulnerable families from accessing critical public support.”
2nd place, Speechmatics track lablab.ai Web Data UNLOCKED, May 2026
A competitor-research dashboard. You enter a company name and, if you have one, an audio clip. It runs web searches and a page scrape through Bright Data, transcribes the audio with Speechmatics, and has an LLM write the findings up as a battlecard.
It placed 2nd in the Speechmatics partner track. The hackathon had 2,310 participants.
Built solo Prototype, May to Aug 2026
A Go prototype for AI agents that have to prove who they are to each other. A small identity provider gives each agent a certificate that lasts one hour and carries a SPIFFE-style ID. Agents then talk over mutual TLS 1.3, a policy engine decides which agent may call which, and messages are signed.
The identity provider doesn't yet check who is asking for a certificate. The fix, which binds each certificate to a known agent, is in an open pull request.
Team lead
UNESCO Youth Hackathon 2026
A fact-checker for health and climate rumours that arrive as voice notes. It checks claims against sources such as WHO, CDC, NASA, IPCC and UNESCO and answers aloud in Urdu, English, Spanish or Indonesian, as a web app and a Discord bot. I led a team of six from Pakistan, Mexico and Indonesia. The event had 2,200 submissions from 129 countries.
Built solo
H0: Hack the Zero Stack, Vercel v0 and AWS Databases
A console for spotting risky local AI processes on company machines. Endpoints send HMAC-signed telemetry to a Next.js app backed by DynamoDB, and an admin can isolate a host in one click. The endpoint file and network events in the demo are simulated.
Built solo
Google Cloud Rapid Agent Hackathon
A venue-operations demo. When the simulated crowd data shows a problem, it pulls context from MongoDB, Elasticsearch and GitLab through MCP, asks Gemini for a plan to reroute people, and shows the plan to an operator to approve.
Live app
IBM Bob Hackathon, May 2026
Paste a Python function and it asks IBM Granite on watsonx.ai for a PyTest suite and a security review, both downloadable. The review is the model's output, not a static analysis, so a person still has to check it.
Playable online
Team entry
AMD Developer Hackathon 2026
A batch agent for a machine with no GPU, 2 vCPUs and 4 GB of RAM. Each task goes first to a 4-bit Qwen 2.5 3B model on the CPU. If the answer fails a check for its task type, the agent asks a Fireworks-hosted model instead.
Built in 120 minutes
In daily use
Personal tool
Every 15 minutes this Apps Script copies coursework from my university and personal Google Classroom accounts into one Drive folder and a Sheets deadline tracker.
Aug 2026 to present
Jun to Sep 2026
Mar 2026 to Dec 2029 (expected)
May 2026
lablab.ai Web Data UNLOCKED, with OmniSight
Jun 2026
USAII Global AI Hackathon, with BISP Benefits Navigator and a 25% certification scholarship
Jun 2026
AI Seekho Google Antigravity Hackathon, Google for Developers
BerriAI/litellm
#47733
BerriAI/litellm
#47858
elizaOS/eliza
7 PRs
Closed after review. Reviewers approved them; upstream shipped its own fixes for the same issues.
Google, 2026: AI Fundamentals, Introduction to Generative AI, Large Language Models, AI Agents and Responsible AI, plus skill badges in Prompt Design in Vertex AI and Build a Website on Google Cloud.
I'm looking for internships and freelance work in AI engineering and application security. Email is the quickest way to reach me.
hamza135252@gmail.com