Chapter I · Groundwork
Training a foundation model on 100+ hours of unlabelled neurosurgical video — learning instrument detection, pose, tissue segmentation, and 3D reconstruction directly from footage.
Enhanced a commercial RAG system for data privacy and automatic user personalisation — training a local LLM with a custom user-embedding layer on RLHF to adapt response style to individual personas.
PCA plot showing the learned embeddings for the 'lawyer' (user_1) and 'social media analyst' (user_2). Their separation demonstrates the model successfully captured their distinct preferences.
I enhanced a commercial RAG system to address data privacy risks and enable automatic user personalisation. I developed a local LLM with a custom user-embedding layer that learns individual preferences from star ratings. The final system adapted its response style to different user personas, making a lawyer persona 90% more likely to receive a detailed, formal response.
The existing system sent user data to third-party APIs, which was unacceptable for clients with sensitive information. The system also had no memory for user preferences between sessions, causing user frustration.
Architecture diagram showing the RAG system, user feedback loop, local LLM, and custom user-embedding layer integration.
My solution uses a Representation Learning Approach. I subclassed a LlamaForCausalLLM to introduce a custom user-embedding layer. This layer concatenates a learned user-specific vector with the standard word embeddings. This method allows a single model to serve all users, which avoids the high cost of training a model for each user.
The model was trained using Reinforcement Learning from Human Feedback (RLHF). I used Group Relative Policy Optimization (GRPO) because its low memory footprint was essential for the single Tesla T4 16GB GPU, as it does not require a reference model like PPO. To make training feasible, I used LoRA, 8-bit precision, and Key-Value (KV) caching, which reduced response times by 1.36 seconds per query.
A projectile is a body that is thrown into the air or projected through space by the action of a propelling force.
A projectile is a body that is shot, flung, or thrown into the air.
Trading agent using CNNs to fuse visual, textual, and trend-based market signals.
Geospatial monitoring system using dual-camera and autonomous drones.
Scalable searchable document system built with independent Docker containers.
ResNet + CLIP ensemble geolocating images to country-level accuracy.
Automated UML diagram generation from source code using static analysis.
Map-based itinerary builder with real-time suggestions and route visualization.
Real-time age estimation from facial images — reliability-focused deep learning.
Ranking system using multiple specialized models for wildlife auctions.
CNN-based monitoring system identifying regions of interest in industrial video.
Retrieval-augmented generation using word embeddings for minimal hallucinations.
LSTM-powered music generation trained on Jazz corpora to create original melodies.
Intelligent trigger word detection using robust background noise training.
Mathematical image processing for calculating cable car wire spans.
Learning to thread a catheter through anatomy with DQN & imitation learning.
Have something to build, train, or scale?
I'm open to research collaborations, freelance engagements, and full-time roles — especially where medical imaging, foundation models, or RL are involved.