Projects

Optimized Bosch's SMT (Surface Mount Technology) assembly line for 32 PCB types with 200+ components.

• Developed algorithm for optimal feeder arrangements improving pick-and-place efficiency
• Increased components-per-hour (CPH) by at least 7%
• Utilized Genetic Algorithms, MILP, and Nonlinear Programming

Research on Tractable Sharpness-Aware Learning of Probabilistic Circuits at Laboratory of Statistical Artificial Intelligence and Machine Learning, IIT Palakkad.

• Collaborated on hybrid training method for PCs - more generalizable than EM, faster than SGD
• Implemented Hessian-based regularizer to reduce overfitting
• Efficient computation of Hessian trace for PCs

An AI assistant handling chat and voice queries that intelligently invokes tools.

• Supports 15+ actions (computer control, real-time weather, web search,etc.)
• Uses Whisper for speech recognition and TTS for voice output
• Modular SaaS architecture with containerized MCP-based tool servers
• Tool routing via LangChain/Ollama, result tracking via MLflow

Semantic search engine for images and videos.

• Uses BLIP for image captioning and Tantivy for indexing
• Deployed dual backends (FastAPI/BentoML)
• Load-tested for 1000+ concurrent requests using Locust
• Ensured clean, configurable microservices with unified logging, ruff, and TOML/YAML config management.

• Conducted rigorous study on Optimal Transport (OT) versus Cosine and Polynomial path schedulers, analyzing their impact on vector field curvature and training stability
• Developed custom visualization suite to render velocity fields and sample evolution on data distributions
• Formulated image denoising as a generative flow problem, adapting pre-trained model to solve inverse tasks by manipulating integration time steps

• Engineered Hybrid Retrieval pipeline over 26,688 Supreme Court judgments, fusing BM25 lexical scoring with dense vector retrieval
• Implemented dynamic thresholding algorithm using elbow detection on relative score drops to auto-truncate ranked lists
• Developed RAG framework with Google Gemini for grounded legal answers, with LLM-based metadata extraction for year/bench filtering

• Extended "Interventional Sum-Product Networks" paper to model causal interventions using custom Graph Convolutional Network (GCN) to encode mutilated causal graphs
• Integrated GCN encoder into iSPN framework to generate parameters for Sum-Product Networks, replacing standard MLPs
• Achieved 26.15% reduction in test NLL on ASIA dataset when compared to a mild improvement on Causal Toy dataset, demonstrating GNN superiority on sparse, large-scale causal graphs

• Objective: Developing a textless speech-to-speech framework to convert neutral synthetic audio into emotion-rich speech (e.g., Joy, Sadness) for mental health and animation applications.
• Data Curation: Addressed the critical scarcity of emotion-labeled datasets by curating a novel corpus from the movie Inside Out, leveraging character archetypes as proxies for emotional states.
• Semi-Supervised Pipeline: Engineered an automated annotation pipeline to resolve missing speaker tags; trained an audio classifier on seed data to iteratively label and expand the dataset using high-confidence predictions.

Slot Booking DBMS for Central Instrumentation Facility at IIT Palakkad.

• Developed comprehensive system with 25+ triggers, functions & procedures
• Role-based access for students, faculty, and staff
• Collaborative team project with intuitive front-end

• Implemented self-supervised denoising pipeline for medical CT scans using Ne2Ne framework, training U-Net directly on noisy data without clean ground truth
• Engineered physics-informed data generation workflow: projecting clean images into Sinograms (Radon Transform), injecting Poisson-Gaussian noise, and reconstructing to simulate authentic scanner artifacts
• Leveraged Ne2Ne's random subsampling strategy to separate signal from noise, restoring diagnostic quality in low-dose CT DICOMs

Context-sensitive information retrieval system using intelligent agents.

• Secured 12th place (out of 23 IITs) at Inter IIT Tech Meet 13.0
• Dynamic thresholding and adaptive retrieval
• Personalized re-ranking for optimal query performance

Predicts stress levels and sleep disorders based on lifestyle and health data.

• Experimented with various ML models with intensive fine-tuning
• Built interactive web app using Streamlit