Projects, Products & Research
Engineering case studies, not product brochures. Each production system below foregrounds a distinct competency — agentic orchestration, document intelligence, retrieval — read as the problem, the decisions, and the trade-offs behind it. Research artifacts, demos, and presentations are linked throughout.
ANA — Audit-Ready Control Testing
Agentic orchestration. A hybrid supervisor/hierarchical system that automates BFSI control testing across 7 steps and 17+ agents, with actor–critic validation in every step and human-in-the-loop decisions.
CovenAce — Covenant Extraction at Scale
Document intelligence. A self-correcting extraction pipeline that pulls financial covenants from legal agreements at 94.47% accuracy — cutting per-batch processing from hours to minutes.
Index AI — Multi-Agent Document Indexing
Ensemble document indexing. OCRs combined-scan banking documents, then runs N agents in parallel to classify and segment each page — resolved by per-page majority voting to 98% accuracy on a synthetic silver dataset.
Pan-Organ Tumor Detection (IIT-B & Tata Memorial)
Collaborative research with IIT Bombay Medical Lab and Tata Memorial Cancer Center. Fine-tuned KimiaNet for multi-organ tumor detection on gigapixel WSIs.
MAP-Boosting
Novel framework treating prompt optimization as pseudo-gradient descent over discrete instruction space using an ensemble of specialized agents.
Fake Emotion Prediction
Research exploring multimodal late-fusion architectures combining text and audio to detect true vs fake internal human emotions.
RAG-Based Code Visualization
RAG system using ChromaDB and Gemini 2.0 to generate complex Manim animation code from natural language mathematical descriptions.
Industrial Fire Detection
Deployed production YOLOv8 fire detection system with real-time video processing and temporal consistency filtering.
SAM Variants & Edge Optimization
Compressed Segment Anything Model (SAM) for edge deployment using ONNX quantization and structured pruning.
UNIFIED: Meta-RL Trading Architecture
Meta-RL framework achieving 10-20x sample efficiency and slashing regime adaptation latency via Transformer World Models.