|
-4169E1?style=for-the-badge&logoColor=white){kind=icon}
An end-to-end AI system that scores SME loan applicants and checks the recommendation against real State Bank of Pakistan regulations before it ever reaches a human. A calibrated XGBoost model predicts default risk and credit limit, SHAP explains every decision, and four LangChain agents handle data extraction, prediction, compliance verification, and final approval — with a hybrid RAG engine making sure no regulation gets cited unless it's actually in the source corpus.
Tech Stack
XGBoost SHAP LangChain Platt Scaling Hybrid RAG (PGVector + TSVector) RRF Reranking
Paste in any lengthy piece of content and get back an LLM-generated summary plus three multiple-choice questions that test whether you actually understood it. Built solo, start to finish, in a 60-minute hackathon.
Tech Stack
Streamlit OpenAI GPT-4o-mini
A real-time attendance system that recognizes faces across multiple camera feeds at once. MTCNN handles detection, DeepFace handles recognition, and Python threading keeps every camera stream running smoothly in parallel — all wrapped in a live PyQt5 dashboard.
Tech Stack
DeepFace MTCNN Multi-threading
| Agents & Orchestration | RAG & Retrieval | ML & Evaluation | Backend & Infra | Languages |
      |  -4169E1?style=flat-square){kind=icon}     |        |       |     |
| diff + ACTIVE → Deploying CreditSense to cloud with CI/CD pipeline + ACTIVE → Building RAG evaluation benchmarks (retrieval precision, answer faithfulness) ! LEARNING → LangGraph for stateful agent workflows with human-in-the-loop ! LEARNING → LoRA fine-tuning for domain-specific LLM adaptation # SHIPPED → CreditSense — 4-agent pipeline + hybrid RAG + ML pipeline # SHIPPED → QuickQuiz-AI — built in 60 min at hackathon # SHIPPED → SFRAS — facial recognition attendance system |
"If it doesn't have an API endpoint, it's not done." — me, probably
