Projects

Civic-tech platform to report water body pollution and track resolution — reached the National Finals of Smart India Hackathon 2017.

• Mobile app for citizens to lodge complaints with photos and track complaint status on a timeline.
• Used YOLOv2 to automatically identify water-related problems in attached photos.
• Web portal for local government staff to view complaints, collate issues by nature and location, and post status updates that reflected in real time on the app.

Final year undergraduate project — built and trained a convolutional neural network to drive an RC car autonomously.

• Collected driving data from a Raspberry Pi camera in manual mode.
• Trained a CNN on steering angle labels; deployed model back on the Pi for inference.
• Car navigated unseen terrain without human input.

Published at the 2023 IEEE Conference on Games. Deep learning models integrating temporal and spatial attention mechanisms for reinforcement learning in ATARI environments.

• Designed attention modules that focused agent perception on task-relevant regions.
• Improved sample efficiency and score across multiple ATARI benchmarks.
• Work completed during M.Tech at IIT Bhubaneswar.

Automated quality scoring for sales and support calls using LLMs and speech-to-text pipelines.

• Increased daily audit throughput by 75%, removing manual bottlenecks.
• Built scoring rubrics configurable per team — sales, academic counselling, compliance.
• Integrated with Twilio for call retrieval and AWS Lambda for async processing.

Retrieval-Augmented Generation chatbot for resolving student queries in real time against a curated knowledge base.

• Used Elasticsearch as the document store with semantic vector search for retrieval.
• LLM synthesised grounded answers — reducing hallucinations compared to vanilla prompting.
• Served via Flask API; integrated into the student portal frontend.

Real-time lead scoring and routing system that cut counsellor contact time from 4 hours to 22 minutes.

• Scored leads using behavioural signals, demographics, and engagement history.
• Routed high-intent leads to available counsellors instantly via queue management.
• Reduced drop-off significantly by closing the gap between lead capture and first contact.

Interactive Voice Response system with dynamic AI-generated prompts to qualify and route inbound leads automatically.

• Dynamic call flows driven by lead profile — different scripts per intent segment.
• Integrated with Twilio for telephony and AWS for compute and storage.
• Reduced dependency on human SDRs for initial qualification calls.

Day-by-day lesson plan tracker that shows what is on time, lagging, or ahead of schedule — and automatically adjusts the plan.

• Tracks every lesson against the planned timeline with live status indicators.
• Supports creating vacations, exams, and custom events via presets or manually.
• Auto-adjusts downstream schedule when changes are made, keeping the plan coherent.