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, user-event context, and automatic DOM screenshots. Automates root-cause diagnosis, hypothesis validation, and SDK-integrated patch synthesis. Evaluated on 69 real-world production bugs from Cal.com and Appsmith, fixing 32/69 issues (vs 9/69 for GPT-4o) and achieved a 72.50% human preference rate across 549 blind A/B evaluations.
FRAC-MAS
A Safe and Explainable Multi-Agent System for Fracture Diagnosis
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). Stacking achieved a 96.4% detection rate on an external dataset.
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). Research paper accepted to IEEE CONECCT 2026.
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.
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.