I collaborated with 3 cross-functional teams to ideate and develop an AI assistant for farmers to help control pest and diseases. This resulted in 20% crop yield for farmers.

I achieved 97% model and evaluation accuracies during the training pest image classification model on 187,858 dataset containing images of plan pest and diseases using Keras

Reduced trained model deployment time to production by 70% using MLFlow, Dagshub and DVC

Collaborated with 3 cross-functional teams to design and develop a highly secured and resilient Authentication Service that made use of token rotation and DPoP.

Maintained latency below 200ms under heavy production workload using SWR, Event-driven architecture, strict clean architecture, CQRS, etc

Led 3 cross-functional teams to design and develop an e-commerce mobile application API endpoints using FastAPI.

Reduced school onboarding cost from ~$2 to ~$0.11 per school by designing a multi-agent extraction system with cost-optimized LLM routing across GPT-4o-mini, Grok 4-fast, DeepSeek, and Gemini 2.5 Flash, replacing what would have been hours of manual data entry per provider.

Cut redundant processing costs from $1 to ~$0.03 by building a resource deduplication layer that allowed agents to reuse scraped or downloaded documents from other agents, eliminating double-fetching across the pipeline.

Achieved 98% extraction accuracy across school details, divisions, grades, schedules, leads, and district data using intelligent model fallback strategies and structured output validation with PydanticAI

Implemented a hybrid scraping stack combining Scrapfly and Browserbase CDP via Playwright, maximizing extraction success (about 98%) across anti-bot-protected and JavaScript-heavy school websites.

Built a reusable agent framework with dependency injection and multi-provider LLM support, enabling rapid development of new extraction agents and reducing agent development time by 70% across the team.

*Evals:* Validated extraction quality at ~95% accuracy using Braintrust with LLM-as-judge scoring against a golden dataset of 400 US schools, ensuring production reliability before rollout.

Reduced kids school public calendar events import cost from ~$1.50 to ~$0.20 per school by architecting a multi-agent pipeline that automatically extracts closures, early dismissals, and notable days — eliminating manual calendar entry for thousands of schools.

Handled 6 distinct school calendar source types (interactive JavaScript calendars with multi-month Playwright navigation, static HTML pages, PDF documents, images via OCR, Google Docs, and ICS feeds), maximizing coverage of 99% across diverse school websites.

Leveraged the shared resource deduplication layer to reuse previously scraped content across ingestion and calendar agents, further reducing per-school processing cost and latency.

Automated end-to-end calendar delivery by persisting extracted events and exporting them to per-child Google secondary calendars with grade-specific filtering, directly reducing parent workload by 90%.

*Evals:* Achieved ~97% extraction accuracy validated via Braintrust with LLM-as-judge scoring against a golden dataset of 200 US schools across all calendar source types.

Achieved 40% improvement in user experience by integrating iOS Live Activities for real-time email ingestion progress tracking using Firebase Cloud Messaging (FCM) v1 API, giving parents instant visibility into processing status.

60% increase in parent engagement by developing scheduled morning and evening digest system delivering personalized event summaries via SMS (Twilio) and WhatsApp (Bird), keeping parents informed without requiring them to open the app.

Reduced school discovery latency by 85% by building a gRPC-based search API with zipcode-to-coordinates geocoding, dynamic radius management, and grid-based spatial caching, delivering sub-second search results for parents finding nearby schools.