[ 03 / Projects ]
Every project here shipped — and each one is measured by what it changed: GPU bills cut, contracts won, hours returned to the team.
Flam
Jun 2025 – Aug 2025
Generative-AI framework that places a human subject into a different background and blends them naturally with the environment. Flux fine-tuned models combined with the IC-Light relighting model generate physically plausible lighting and shadows; image-processing passes smooth the borders. Delivered one week ahead of schedule.
≈35% lower GPU cost via inference and GPU management
Flam
Mar 2025 – May 2025
Wall detection and paint visualisation pipeline pairing YOLOv11 segmentation with a fine-tuned Stable Diffusion model. Renders the chosen paint colour onto walls while preserving light reflections on reflective surfaces. Deployed as a containerised inference service.
Proof of concept secured a client contract worth ≈£42,500
Tata Elxsi
Jul 2024 – Feb 2025
Multi-agent framework built on OpenAI GPT models that captures user inputs, updates project progress and generates client reports. NeMo Guardrails blocks prompt attacks; a Qdrant vector store powers retrieval (RAG). Later adapted for the IT department's ticketing system.
≈60% reduction in reporting and analysis time
Tata Elxsi
Apr 2023 – Apr 2024
RAG system converting high-level human instructions into executable robotic behaviour trees with GPT-3.5-Turbo, grounded on robotic-arm data and available AI skills via a Weaviate knowledge base. Owned prompt engineering, the RAG pipeline and productionisation.
Demo shipped in 3 weeks; showcased at iREX 2023, Tokyo
Tata Elxsi
Apr 2024 – Jun 2024
Self-ideated pipeline generating datasets of defective industrial components from 3D CAD models rendered with Pyrender, with InstructPix2Pix modifying textures to simulate cracks and rust. Deployed internally with Docker and FastAPI.
≈19% lower project costs; customer signed a SaaS subscription proposal
Tata Elxsi
Nov 2021 – Apr 2023
Systematic research and fine-tuning of neural pose-estimation models for reflective, textureless industrial bolts. Built an OpenCV annotation tool to unblock dataset creation, later semi-automated. Full research process documented and delivered to the Japanese customer.
≈80% reduction in manual annotation time
Undergraduate research
2020 – 2021
Medical-imaging segmentation model optimised to run on a 4 GB Jetson Nano edge device — final-year undergraduate project demonstrating deployable deep learning under tight hardware constraints.
IEEE National Level Best Project Award 2021