Mahiyadav Sidda

About

Built an "Offline Study Assistant" that runs LLM inference locally on a 5-year-old Android device using Llama.cpp and the Android NDK, achieving a 27x speedup and cutting time-to-first-token from 11 minutes to 30 seconds. Also has applied backend/API experience with FastAPI, Supabase (Auth + RLS), and production hardening of a RAG system at Hashmint using Celery and Redis to eliminate PDF-processing-related query failures.

Experience

Education

Key Strengths

• Optimized on-device LLM inference from 11 minutes to 30 seconds time-to-first-token (27x faster overall) on Snapdragon 720G
• Designing resource-constrained inference pipelines with focus on latency, memory, and battery tradeoffs
• Scaling FastAPI services with Gunicorn multi-workers and Redis caching; enabling frontend via CORS and OpenAPI
• Implementing API security controls (rate limiting, API keys) for ML endpoints
• Managing migrations/refactors with transactions, validation hooks, blue-green cutovers, and canary rollouts (10% traffic) using Supabase connection pooling
• Improved RAG system robustness by addressing PDF chunking/doc-processing bottleneck; reduced ~15% query failures using Celery + Redis queues