Featured Projects
Selected work spanning multimodal hate-speech detection, civic-tech analytics, and computer-vision healthcare applied with end-to-end product thinking.
AntiBug
Runtime-Aware Multimodal Agentic Program Repair
Architected a 12-agent LangGraph pipeline that ingests source-map-resolved stack traces and automatic DOM screenshots. Automates root-cause diagnosis, hypothesis validation, and SDK-integrated patch synthesis. Evaluated on real-world production bugs, fixing 46.38% of issues (3.5x better than GPT-4o base).
FRAC-MAS
Multi-Agent Fracture Radiograph Pipeline
Engineered a clinical decision support system integrating a 4-model vision ensemble (MaxViT, RAD-DINO, YOLO, DenseNet) with specialized LLM agents. Implemented conformal prediction to generate statistically grounded differential diagnoses for X-rays (92.0% empirical coverage, 96.4% detection rate).
DigniFy
Multi-Modal Multilingual Hate Speech Detection
Engineered a mobile application pipeline with DistilRoBERTa for text, OCR for image extraction, and STT-TTS for audio. Features an AI Agent integrating all modality-specific models as tools, orchestrating interaction across four LLMs (DeepSeek R1, GPT-4o, Gemini 2.0, LLaMA 4.0).
MetaSearch
Search-Augmented LLM for Consensus Resolution
Designed a reasoning-based AI Agent combining disagreement detection and fact-grounded synthesis to automate meta-review decision-making in peer review. Utilises specialized heuristic constraints based on Cognitive Load theory to optimize consensus generation.
Glass House Argumentation
LLM-Argumentation System for Peer Review
Developed a novel LLM-based argumentation system evaluating 900+ ICLR reviews using leading LLMs (Llama 3.1, Mixtral, Gemini 2.0, GPT-4o) to enhance interpretability of peer reviews, achieving high inter-annotator agreement (Cohen's Kappa = 0.834).
RakshakRita
QR-led Citizen Feedback System
Civic-tech solution for police departments with interactive geo-spatial dashboards, multilingual sentiment analysis, and automated report generation. Recognized as a Top 100 finalist at the Rajasthan Police Hackathon 2024.