Applications & Projects

An open-source framework for evaluating medical AI systems in Indian clinical contexts. Includes four domain-specific evaluation datasets covering ASR, document understanding, clinical NLP, and multilingual tasks. Built because most medical AI benchmarks don't reflect real Indian conditions.

Open-source release of EkaCare's Parrotlet model family on HuggingFace — medical ASR (Parrotlet-A), vision LLMs for medical documents (Parrotlet-V), and multilingual medical embeddings (Parrotlet-E). Purpose-built for Indian healthcare, available for research and clinical use.

Open evaluation datasets for medical AI in Indian healthcare contexts — covering ASR benchmarks, document understanding, multilingual clinical NLP, and structured extraction tasks. Released alongside KARMA and IndicMTEB to enable reproducible research.

A comprehensive benchmark for evaluating multilingual medical text embeddings across 22 Indian languages. Built to fill the gap in Indian healthcare NLP evaluation — MTEB and similar benchmarks are English-first and clinically naive.

Leading an industry initiative to build a digital public good for seamless AI scribe integration with EMR systems across Indian healthcare. Defines an open, vendor-agnostic protocol covering handshake, FHIR-compliant data exchange, and translation adapters — so any scribe vendor can plug into any EMR without proprietary lock-in.

Bharat Ontology for Disease & Healthcare Informatics — open-source clinical knowledge graphs for grounding healthcare AI in verified clinical data. Covers 779 conditions, 4,037 symptom variants, 1,186 drugs, 812 lab tests, and 39 specialties. SNOMED-linked, production-tested. Available in Neo4j, CSV, PyTorch Geometric, and RDF formats.

Clinical AI assistant combining voice-powered documentation, intelligent patient profile management, and real-time clinical insights. Features drug interaction alerts, multi-language support, and automated SOAP note generation.

Measures heart rate using the smartphone's back camera via photoplethysmography — no additional equipment needed. Useful for fitness tracking and detecting when medical consultation may be necessary.

Large-scale open dataset for research on Carnatic and Hindustani music, featuring multitrack audio, melodic annotations, and metadata. Published in Empirical Musicology Review 2021.

AI-powered practice app for Indian classical music providing real-time feedback on pitch accuracy, tala adherence, and raga grammar. Co-founded MusicMuni Labs to build and commercialize this. 1M+ installs.

Doctoral research on computational approaches for melodic description in Indian classical music corpora — melodic similarity, pattern discovery, raga recognition, and tonic identification. CompMusic / ERC project at MTG, UPF.

European Research Council funded project at MTG, UPF for computational analysis of non-Western music traditions. Contributed tonic identification, raga recognition, melodic similarity, and corpus tools for Hindustani and Carnatic music.

Open datasets released alongside my PhD thesis — annotated melodic data for Hindustani and Carnatic music, including tonic labels, raga annotations, and melodic similarity judgments. Widely used in subsequent MIR research on Indian art music.

Web platform for exploring and analyzing the CompMusic music corpora — large collections of annotated audio recordings across Hindustani, Carnatic, Makam, Beijing opera, and Andalusian musical traditions. Built at MTG, UPF.

Real-time raga recognition from audio using pitch class profiles and machine learning. Identifies the raga being performed and visualizes melodic patterns live. Built at HAMR 2015.

A musical game inspired by the Sawaal-Jawaab tradition in Indian classical music. Players sing or play a melodic phrase and the system responds with a melodic answer. Built at HAMR 2015.

Replaces all speech in any audio/video with Hodor's iconic vocabulary while preserving the original prosody, rhythm, and emotional tone using voice conversion techniques.

Essentia-powered implementation of the Infinite Jukebox concept — analyzes a song's beat structure to find perceptually similar segments and creates an endlessly recombining loop that always sounds musical.

Python library developed as part of the CompMusic project at MTG, UPF — tools for accessing and processing the CompMusic music corpora, including audio feature extraction, metadata handling, and integration with the Dunya platform.

Open-source software for sound analysis, transformation, and synthesis using spectral modeling. Contributed as part of the ASPMA Coursera MOOC at MTG, UPF — the course taught signal processing fundamentals to 100K+ students worldwide.

JavaScript implementation of the YIN pitch detection algorithm for real-time fundamental frequency estimation — pitch tracking directly in the browser, no server needed.

Contributed to Essentia during my PhD at MTG, UPF — an open-source C++/Python library for audio and music analysis. Now a standard tool in music information retrieval research, used widely across academia and industry.

Transforms Western pop songs into Hindustani classical style by applying raga-based melodic transformations — converting Western scales to raga scales and adding characteristic ornaments. Built at EchoNest Hack Day 2013.