Things I've
designed, built
& shipped.

A 155-class crop-disease classification framework spanning 21 crops and 298K+ images. A crop-aware routing system assigns each image to a dedicated CNN expert, with SCOLD integrated as a confidence-gated vision-language expert for multimodal reasoning. Benchmarked fusion and meta-learning strategies against ConvNeXt, EfficientNet, ResNet50 and SCOLD. Earned a 4.0 GPA and a university Open Day poster.

A fully air-gapped retrieval-augmented generation system for confidential document retrieval and offline answer generation — on-prem indexing, zero external calls, telemetry removed. A production-grade ingestion pipeline handles mixed formats with incremental upserts, deduped chunk IDs and deletion sync. Answer faithfulness improved through hybrid retrieval, cross-encoder reranking and multi-hop query decomposition.

A peer-tutoring marketplace connecting students and tutors through a free online platform. Built from scratch with Agile planning — weekly meetings, roadmaps, issues and user-focused milestones. Led a team of five engineers across UML diagrams, JavaDocs, storyboards and maintainable application code, designed around usability, efficiency, security and API-backed functionality.

A Streamlit chatbot with persistent history, PDF upload, and Plain / RAG-Fusion / Router-based RAG pipelines — implemented end-to-end from scratch. An LLM critic scores and compares variants in a single run. Also includes a tool-calling agent (web search, task management, device state) and a multi-agent deliberation pipeline where specialist agents propose alternatives and an orchestrator produces the final response.

Led a team of five in building a sentiment-analysis system for Urdu news articles, from data collection to model evaluation. Developed custom web-scraping pipelines to extract Urdu-language news from multiple sources, conducted EDA to uncover linguistic patterns, and trained a classifier to categorize content by emotional tone and topical relevance.