Most AI demos fail the same way: they answer confidently, they answer wrong, and you only find out afterwards. That is tolerable in a chatbot and unacceptable in a system someone acts on a government procedure, a credit decision, a bias audit.
So the question I keep circling is the boundary: which parts of a system are allowed to be a language model, and which parts must be code you can test. Most of what I build sits on that line.
- 🔭 Building Government Service AI Navigator live, deployed, and evaluated against 51 scripted scenarios
- 🧪 I never claim "grounded" without a number attached there is a harness, and CI fails the build if one unsupported claim gets through
- 👁️ Day to day: real-time computer vision YOLO, tracking, depth and distance estimation
- 🗣️ Ships in English, Urdu and Roman Urdu
- 💬 Ask me about grounded RAG, three-valued logic, or why a good fallback beats a confident guess
Assistant Manager AI · Stingray Tec · 2026 Present
Production AI systems computer vision, generative AI and LLM applications built to run in the field.
Real-time video with custom-trained YOLO models, object detection and tracking; monocular distance and
depth estimation for perception; RAG pipelines and agentic workflows served through Python and FastAPI.
Full lifecycle: data → model → inference → integration → deployment.
Research Assistant Generative AI in Education · FAST-NUCES · Summer 2025
Researched where generative AI genuinely helps in higher education LLM-based tools for teaching,
assessment, knowledge assistance and student productivity.
Full Stack Developer Intern · XOLVA · Summer 2023
Full-stack applications with React.js, Node.js, Express and PostgreSQL. REST APIs, backend services,
database architecture, and frontend/backend integration.
A Pakistani citizen types "mera CNIC gum hogya hai, Karachi mein hun" and gets back a personalised, source-cited action plan: which service, which branch, which documents they specifically still need, which office, what happens next.
It is built around one rule:
Not a fee. Not a deadline. Not a document name. Every such fact traces to a database row or a retrieved chunk of an official document, and carries its source to the screen. The model detects intent, phrases questions, and translates nothing else.
The test for whether that boundary actually holds: swap the model for a template renderer are the
answers still correct? They are. LLM_PROVIDER=mock runs the whole system that way, and the evaluation
suite still passes.
| 🎯 Interview | Asks only questions whose answer can change the outcome real information gain over the rule set, not a shortened form. Averages 4.5 questions. |
| 🔍 Grounding | An output verifier scans every rendered number, duration, count and URL, and rejects anything not traceable to a fact including numbers that drift during translation. |
| 📊 Evaluation | 51 scripted citizen paths · 3 services · 3 languages · prompt injections included. 100% service and scenario identification, 100% document F1, 0 unsupported claims. |
| 🤲 Honesty | Unverified fees ship as NULL and render as "not verified confirm at the counter." A plausible invented number is worse than a blank one. |
AI / ML
Computer Vision
Languages
Backend & Data
Frontend & Tooling
| Project | What it is | Stack |
|---|---|---|
| 🏛️ gov-service-navigator | Grounded citizen-services decision engine. Live, tested, deployed. | Next.js pgvector RAG |
| 🗺️ Manzil | Multi-agent trip planning. I built the agent orchestrator and the evaluation harness. | Python FastAPI RAG |
| 🤖 Autonomous-Enterprise-OS | Multi-agent orchestration platform for enterprise workflows. | Python FastAPI LLM |
| ⚖️ Bias-Detection | Detecting bias in text the measurement side of responsible AI. | PyTorch Transformers |
| 💳 Credit_Card_Default | Default prediction on the Taiwan credit dataset. | scikit-learn Jupyter |
| 📊 IBA_Datathon | Competition work under time pressure. | Python Pandas |
| 🔐 Escrow | Escrow payment system on the PERN stack. | PostgreSQL React |
Coursework lives in Compiler-Construction, Information-Security and DevOps kept public because the working is worth more than the grade.
interfaces for people who are anxious, in a hurry, and on a cheap phone.