Career Kompass: AI Career Coach MVP
Built and deployed an MVP for the founder of Career Kompass. Users paste a job description and upload a CV to get a per-skill gap analysis with match scores, a JD-tailored CV rewrite with before/after diff, a cover letter, and a prioritised action plan. FastAPI backend using DeepSeek with automatic fallback to Llama 3.3 70B on Groq, Supabase auth and saved history, and CV export to PDF and DOCX in multiple templates. CV files are processed in memory and never stored (GDPR by design). Deployed on Vercel and Render.
Generative AI for Aerial Image Analysis
Master's thesis. Evaluated open-source VLMs (Qwen-2.5-VL-7B) via zero-shot & few-shot prompting to classify attic usability across 745 annotated aerial building images, achieving up to 81.6% accuracy. Built a multimodal inference pipeline linking imagery to PostgreSQL geospatial records. Benchmarked 6 VLMs and analysed XAI failure modes including hallucination and spatial misalignment.
GenReq: AI for Requirements Engineering
Co-led architecture of a full-stack AI tool reducing SRS generation from days to 20–30 min at ~$0.60/session. Built a RAG pipeline with LangChain, ChromaDB, and FastAPI for auto-generation of requirements specs, UML diagrams, and LaTeX-compiled PDF reports. Led architecture evolution from Streamlit/SQLite to a React + MongoDB production stack.
XAI for Graph Neural Networks
Analysed the heterogeneous MovieLens100K graph (943 users, 1,682 movies, 80K edges) in PyTorch Geometric. Applied Captum Integrated Gradients to explain a SAGEConv-based GNN via link prediction. Ran ablation experiments (feature-zeroing and timing-edge removal) to validate XAI behaviour on categorical graph data.
AR/XR Chemistry Lab: ARChemLab
Developed intuitive UI features for an augmented reality chemistry lab application using Unity. Deployed the final solution on Microsoft HoloLens 2 to support immersive, interactive digital learning experiences in higher education.
Student Exam Performance Predictor
ML web application built with Flask to predict a student's Math score based on demographic and academic inputs. End-to-end pipeline covering EDA, feature engineering, model selection with Scikit-learn, and web deployment via Flask.
