Where I've worked

Founded Pulp AI to help product teams pre-screen A/B variants against up to 1,000 AI-generated customer personas before committing to a live test.

• ▪Shipped the full-stack MVP in 3 weeks on React, FastAPI, Postgres with pgvector, Redis, and Docker with end-to-end CI/CD
• ▪Built a multi-model routing engine across GPT families to balance cost and latency per task. Personas were sampled from a demographic prior, each with a unique prompt, and the resulting reasonings were clustered to surface dominant cohorts
• ▪Back-tested against 10 historical A/B tests, predicting the winning variant with 70% accuracy against a 50% random baseline
• ▪Owned GTM from ICP definition and pricing through waitlist-to-beta conversion and design-partner onboarding

Founded Edvo, an adaptive assessment co-pilot built in Flutter that closes each student's mastery gaps in real time using an LLM judge loop running over a skill graph.

• ▪Each next question is selected by an LLM reasoning over the student's history, drawing from a hand-curated question bank and generating on the fly when no suitable match exists
• ▪Field-piloted across after-school learning centres over one month and 10 tests, with students showing measurable gains in time-to-mastery and retention

Validated a peer-to-peer rental marketplace for high-value scuba gear across Indian dive communities, running a no-code MVP before committing to a full build.

• ▪Stood up the MVP using a landing page, WhatsApp for coordination, and UPI for payments, keeping build cost near zero while capturing real demand signals
• ▪Tested product-market fit with live transactions and user interviews across India's scuba diving communities

Led two national-scale ML deployments at India's NCDC and across thousands of farms, covering real-time disease surveillance and agricultural pest detection.

• ▪NCDC Outbreak Alert Pipeline. Built a multi-stage, multi-lingual detection system combining rule-based filters, BERT classifiers, LLMs, and clustering to scan 400M+ media sources, issue 100K+ alerts, and verify 5K+ outbreaks. Published at ACL 2025
• ▪Pest Detection. Trained ResNet, Inception, and EfficientNet models to identify pests on cotton and rice farms. Advanced augmentation improved robustness by 14%, directly helping 10,000+ farmers

Deployed ML systems for TB treatment adherence prediction and oral cancer screening, both at national scale in partnership with India's public health infrastructure.

• ▪TB Adherence Prediction. Trained a similarity-encoded ensemble of XGBoost, LightGBM, and EBM on 0.8M Nikshay TB cases, reaching AUC-ROC 0.80 and Recall@20 of 0.62, roughly 3x the rule-based baseline. Deployed across 15 states and flagged 100K+ high-risk patients. Best Paper at ML4H at NeurIPS 2022
• ▪Oral Cancer Screening. Built a lightweight CNN for early-stage oral cancer detection. Diffusion-augmented training lifted recall from 79% to 91%. Co-led data collection and labelling with PGIMER Chandigarh across 10K+ oral lesion images

Built probabilistic record-linkage systems for India's National Digital Health Mission and the CoWIN vaccination platform.

• ▪Health ID De-duplication. Built a de-duplication engine combining TF-IDF blocking, MinHash-LSH, and GBDT to merge 90M+ patient records at 96% precision and 92% recall, cutting manual review by roughly 80%
• ▪Adapted the same pipeline for CoWIN to screen sign-ups and block duplicate vaccination attempts during India's nationwide COVID vaccination drives