All Projects

Final year project. Trained a BiomedCLIP modality router on 14k images reaching 100% test accuracy, plus SMP UNet++ segmentation models hitting 95% accuracy and 88.5% Dice across 7 imaging modalities. Models published to Hugging Face behind a full-stack Django platform with REST endpoints and PostgreSQL-backed patient data pipelines.

PyPI-published agentic memory library with hierarchical concept graphs, BFS traversal, and hybrid RRF retrieval, with 179 tests, fully async, and adopted by external developers. Ships a FastAPI service and React dashboard.

Open-source, privacy-first AI screen & meeting copilot. Transparent always-on-top overlay invisible to screen share, with real-time OCR and audio transcription.

AI-assisted smart contract security platform that detects vulnerabilities, validates exploitability with reproducible PoCs, and writes verifiable audit attestations on-chain bound to the exact audited code.

Open-source observability platform for multi-agent AI systems. See exactly what your agents are doing: every step, every decision, every cost.

LLM-powered bilingual (Urdu/English) clinical web app using LangGraph multi-agent orchestration across 5 modular AI services for patient triage, symptom analysis, and report generation. Structured output parsing and entity extraction with PostgreSQL storage behind a clean FastAPI backend.

Production face recognition SDK combining RetinaFace detection, MagFace embeddings, and Qdrant vector search for fast identity matching at scale.

Explainable quantum-enhanced DenseNet121 for monkeypox skin lesion classification, research combining quantum circuits with deep learning for medical image analysis.

Full-stack legal document ingestion and query platform with a FastAPI backend and a React (Vite) frontend.

Full-stack SaaS that generates complete YouTube Shorts from a URL or idea, producing a 9:16 MP4, SEO description, hashtags, and thumbnail.

Django web app integrating ArXiv, Google Scholar, PubMed, and IEEE for paper search and summarization via LangChain and the Groq API.

AI-powered social media post generation and scheduling tool.

End-to-end framework for recording, replaying, and validating Computer Use Agent trajectories, building SFT-ready datasets from file, scripting, API, and browser workflows.

QLoRA fine-tuning pipeline for natural text generation that pulls human-written text from 5 Hugging Face datasets and trains a LoRA adapter on Llama 3.2 3B.

End-to-end SFT + DPO training pipeline for humanizing AI-generated text, with LoRA checkpoints and GGUF export for local inference.

Multi-component AI text detection system inspired by GPTZero, combining GPT-2 perplexity, burstiness scoring, and a fine-tuned DeBERTa-v3 / Qwen2.5-3B classifier.

Explainability framework reaching 98.4% binary and 86.8% 12-class accuracy using RoBERTa embeddings with stylometric features, LIME, SHAP, and Captum.

Scalable RAG pipeline with multi-source document ingestion (PDF, DOCX, HTML, images) and semantic search over vector embeddings, combining dense and sparse hybrid retrieval. Tuned with chunking strategies, re-ranking, and query augmentation, containerized for reproducible ML-ready deployment.

Automatic Python docstring generation with CodeT5, shipped with a REST API, demo interface, and model evaluation harness.

Production-grade, privacy-first counseling assistant with three layers of safety guardrails and epistemic humility controls.

Mental health mobile app with a personalized AI chatbot. LLaMA 2 fine-tuned on therapist-client data and quantized for efficient mobile deployment.

Vision Mamba + UNETR architecture for retinal vessel segmentation in medical imaging.

AI-powered CPAP mask fitting using DECA 3D face reconstruction to recommend the right mask geometry per patient.

Active learning loop over MagFace embeddings to cut labeling cost in face recognition datasets.

PyTorch ViT implementation with Tiny/Small/Base variants, Mixup/CutMix augmentation, and LR scheduling on CIFAR-100.

YOLOv11 face detection with EfficientNet V2 classification, MTCNN preprocessing, and a real-time pipeline.

FinBERT-powered financial sentiment pipeline ingesting 500+ articles per day from 18 sources, with a multi-signal weighted macro scoring system (DXY, yields, oil, VIX) and entity-level aggregation. Real-time news feeds a live gold and silver calculation surfaced on a Streamlit dashboard.

GPU kernel programming challenge. Writing CUDA kernels from scratch to build deep understanding of GPU computing.

Training agents to play Flappy Bird using the NEAT evolutionary algorithm against a reinforcement learning baseline.

Full compiler across all six phases: lexing, parsing, semantic analysis, IR generation, optimization, and x86-64 assembly output. 4000+ LOC.

PyTorch Transformer built from first principles: scaled dot-product attention, multi-head attention, positional encoding, and the full encoder-decoder stack.

OOP semester project covering full banking operations. 32 stars on GitHub.

Data structures and algorithms reference implementations in C++, used by students across the university.