Multimodal Document Understanding VLM

LLaVA-1.5 7B QLoRA PEFT FastAPI Computer Vision

• Fine-tuned LLaVA-1.5 on document-image QA — +12% over zero-shot on DocVQA
• OCR-free tokenisation + perceptual-hash dedup pipeline
• Constrained JSON decoding for RAG integration; served via FastAPI

6-DoF Pose Estimation — Industrial Bin Picking

PyTorch TensorRT ONNX OpenCV PnP/RANSAC

• 22% latency reduction via TensorRT INT8 — ADD/ADI metrics preserved
• ONNX → TensorRT engine; async batched multi-camera inference
• Deployed at L&T LTTS in Dockerised REST service

RAG Document QA Pipeline

LangChain ChromaDB FastAPI AWS EC2

• End-to-end RAG: PDF ingestion → chunking → dense embeddings → LLM synthesis
• Hybrid retrieval (BM25 + dense) for improved recall on technical docs
• Streaming REST API; deployed on AWS EC2

Neuro-Symbolic Safe Agent (Cognitive AI)

PyTorch LLM VAD Bottleneck Cognitive Modelling Docker

• Cognitive architecture: separates neural affect perception from deterministic policy
• Stress-tested against 100+ adversarial prompts via VAD semantic bottleneck
• Suppresses unsafe instruction-following; Dockerised REST deployment

Chagas Disease Detection — PhysioNet 2025

PyTorch CNN–LSTM WFDB scikit-learn

• 12-lead ECG classifier across CODE-15%, SaMi-Trop, PTB-XL datasets
• Platt probability calibration + AUROC harness; regression tracking

EEG Foundation Model (Cognitive Neuroscience × ML)

PyTorch Transformers Contrastive Learning Neuro-Signal Processing

• Contrastive pre-training for EEG embeddings with cross-subject transfer
• MAE-style masking adapted for cognitive time-series signals