Projects
Vortex
A real-time wealth management system achieving 31.9% annual returns, combining AutoEncoder-based risk scoring, clustering, streaming ML, inverse reinforcement learning for portfolio optimization, and vector-based indexing with explainable AI.
Time Series Analysis
Financial time series suffer from high noise, making direct price prediction difficult. This framework uses deep learning architectures to forecast key structural dynamics such as volatility regimes and momentum states to generate high-conviction trading signals across asset classes. It pairs these signals with fee accounting, market friction modeling, and adaptive risk management to optimize execution timing and maximize risk-adjusted returns.
Physics-Guided Neural ODE for COVID-19
Developed a physics guided neural network which models COVID-19 dynamics across India using population infection data over a 3 year period. A neural network learns time-varying transmission, incubation, recovery, and mortality parameters from epidemiological and mobility-related features. Integrated the SEIRD equations using a differentiable RK4 solver, enabling end-to-end backpropagation through the epidemic dynamics.
OrganTrack
3D medical scans suffer from slice-to-slice boundary ambiguity and high annotation costs. This framework leverages video tracking architectures and memory attention to propagate sparse 2D prompts into spatially coherent 3D volumetric segmentations, combining foundation models with parameter-efficient fine-tuning for accurate clinical analysis.
Graft
GRAFT is a local-first GraphRAG system that combines hierarchical knowledge graphs, RAFT fine-tuning, 4-bit QLoRA, and LangGraph-based retrieval to enable efficient multi-document algorithmic reasoning on low-VRAM hardware.
CreekUI
CreekUI is a mobile-first AI image editing system combining semantic image analysis, parametric style representation, and sketch-to-image generation. It uses Flutter, on-device ML, quantized vision models, and privacy-aware generative backends for low-latency editing.
Causal Extraction
A graph-based causal analytics pipeline for root-cause analysis across customer-agent transcripts. Builds Dialog2Flow conversation graphs in Memgraph and traces causal chains through them to surface the real drivers behind customer escalations.
Visual QA
A multimodal AI assistant that understands images, describes scenes, reads text, answers visual questions, handles conversational follow-ups, and uses uploaded documents for context. It combines a vision encoder with Qwen2.5-3B to generate context-aware responses.
NLP Clarification
When you give an AI a coding task that isn't clear, it usually just guesses what you meant, and if it guesses wrong, the code is wrong. This project fixes that by letting the model ask a few simple yes/no questions first, written as small unit tests, before it writes the final code. Based on the paper "20 Questions for Code," it uses Expected Information Gain to pick the most useful question and stop asking once confident. Tested two approaches, a small model that writes its own questions and a bigger 7B model, on the MBPP and MBPP+ datasets.
CCTV
A multi-stage person tracking system combining YOLO11, part-based ResNet50 embeddings, Kalman filtering, Hungarian assignment, and persistent ReID galleries for robust identity tracking through occlusions, fast motion, and long-term re-entry.
Agentic Research Paper Retrieval
An Agentic RAG pipeline that automatically researches a competition problem, discovers and filters papers, indexes them, and generates a cited literature review. It combines LLM-based planning, arXiv and Semantic Scholar search, citation chasing, hybrid retrieval, reranking, and verification to reduce irrelevant results and hallucinations. A 70B LLM handles critical grading and generation, while bounded retries and calibrated refusals keep unsupported claims out of the final report.
3D Cloud
This project turns a single photo into a 3D point cloud, with no training needed. It finds the object, cuts it out, guesses its depth, and then places each point in 3D space with its color, ready to view in any 3D viewer.