Passionate about AI/ML with both theoretical understanding and Applied AI knowledge through projects. Backed by solid experience in full stack mobile/web development and deployment.
AI/ML: Generative AI Systems, LLMs, Agentic AI (voice, multi-agent, tool-based), RAG, Fine-Tuning (CPT, SFT, RLHF, RLVR), LoRA/QLoRA, RL Envs, Verifiers, Harness, Evals, Memory/Knowledge Systems, LangGraph, PyTorch, Tensorflow
Mobile/Web: Flutter, React Native, React, Node.js, FastAPI
Languages: Python, JavaScript, TypeScript, Dart, C++, SQL, Bash
Databases: PostgreSQL, MongoDB, Firebase, Supabase, MySQL, ChromaDB
- AI Engineer @ BitBop (Aug 2026 -- Present, Islamabad, PK) — Architecting and delivering production-grade AI systems for education technology — including intelligent timetable generation, LMS automation, and AI-driven exam creation workflows — owning the full pipeline from LLM integration and agent design through system optimization and deployment
- AI Intern @ App In Snap (June 2026 -- July 2026, Islamabad, PK) — Architected and shipped production AI systems including a voice banking assistant and a multi-platform social media automation product, owning the full pipeline from LLM integration and agent design through infrastructure optimization and deployment
- React Native Developer @ Tulip Tea (June 2025 -- Sept. 2025, Islamabad, PK)
- Flutter & Node.js Developer @ 92 Devs (June 2024 -- Sept. 2024, Islamabad, PK)
- AXON -- Voice Banking Assistant (FastAPI, LangGraph, Groq, STT/TTS, RoBERTa) — Built a full voice pipeline for Bank Islami: WebSocket audio stream → STT → router agent → 14 specialized action agents or chat agent, with structured JSON throughout; replaced the LLM-based router with a fine-tuned RoBERTa intent classifier and replaced LLM action agents with NER-driven entity extraction (GLiNER) to cut latency and infrastructure cost while maintaining accuracy
- SupportPilot -- Multi-Channel Social Media Automation Platform (FastAPI, LangGraph, PostgreSQL, Meta APIs) — Centralized automation hub supporting WhatsApp, Instagram, and Facebook with multi-account management; admins can create AI agents and assign them to DMs or comment threads, configure keyword-triggered response templates, and manage all channels from a single workspace; architected a multi-tenant PostgreSQL schema supporting per-account permissions and cross-channel agent assignment; planned extension to call automation via voice agent
- Attention Is All You Need (PyTorch) — Implemented the full Transformer pipeline from scratch: tokenization, positional encoding, and Q/K/V self-attention
- MiniLLM (PyTorch) — Trained a GPT-style LLM on Shakespeare featuring BPE, RoPE, KV cache, GQA and SwiGLU activations
- Nexus -- Autonomous Development Agent (Python, OpenRouter) — Multi-agent system with a dedicated coding agent, research agent, ReAct reasoning loop, context summarization for long task horizons, and a plan-then-execute module for structured decomposition of complex problems
- RAG Pipeline (Python, OpenRouter) — Hybrid semantic + BM25 search, LLM-based chunking, contextual enrichment, and cross-encoder reranking
- Recursive Language Model (Python, OpenRouter) — LLM scaffolded inside a Python REPL enabling code execution and dynamic sub-agent spawning; based on the RLM paper by Zhang & Khattab
- FinSight -- Financial Domain Adaptation (Fine-Tuning) (Unsloth, LoRA, QLoRA, Hugging Face)
- CPT — Domain-adapted SmolLM-135M on SEC 10-K filings using LoRA/QLoRA; resolved catastrophic forgetting by mixing general and financial data during training
- SFT — Shaped instruction-following behavior using Alpaca prompt template on financial-qa-10k; merged CPT adapter before layering SFT; applied sequence packing, loss masking, and EOS tokens
- RLHF — Applied PPO-based optimization using TRL to align model outputs with human preference; trained a Reward Model as a proxy judge on chosen/rejected pairs; optimized policy against a frozen SFT reference model with KL penalty to stabilize updates and prevent reward hacking
- RLVR — Applied GRPO with rule-based verifiable rewards (exact-match on financial QA), eliminating the learned Reward Model; used group-relative advantage estimation to stabilize policy updates
- Final Year Project (Blockchain, AI, React, Node.js, PostgreSQL) — Ethereum-based medical EHR system with patient-controlled data access and an integrated AI medical coding module
- MovieFlix (React Native, Expo) — Movie discovery app with TMDB API integration, Expo routing, and custom UI
-
Healthy Ways (Flutter, Supabase) [Live at Play Store
$|$ App Store] — Role-based medication reminder app for patients, doctors, and pharmacists - Skribbl Clone (Flutter, Node.js) — Real-time multiplayer drawing game with WebSocket sync and in-game chat
- Islamic Application (Flutter) — Feature-rich Islamic utility app with prayer times, Quran reader, Qibla compass, and Tasbih counter
| Year | Degree | Institution |
|---|---|---|
| 2022 – 2026 | BSc Computer Science | Air University, Islamabad |
| 2020 – 2022 | Intermediate (ICS) | Fazia Inter College, Kallar Kahar |
| 2018 – 2020 | Matric (Computer Science) | Fazia Inter College, Kallar Kahar |
