MANIK RANA

Designed and implemented Puch's core AI agent loop and LLM harness using Python/FastAPI, incorporating advanced tool orchestration and context management capabilities.

Refined system prompts and conversation behaviors, successfully increasing session engagement by a notable 60%.

Co-designed a RAG document agent utilizing hybrid BM25/vector retrieval, which accelerated data ingestion by 300% and reduced context pollution by 50%.

Developed browser- and mobile-use tooling for AI agents, enabling navigation of websites and control of Android apps with secure user handoff for authentication and payment.

Engineered end-to-end web and React Native applications, supporting streaming chat, citations, and advanced image/video generation features.

Built analytics and feature-flag services using PostHog, ingesting and backfilling sessions for over 1,000,000 users via scheduled jobs and Redis checkpoints.

Led the comprehensive frontend redesign for TurboML using Next.js, React Server Components, and TanStack Query caching, reducing load time by 40% and redundant requests by 35%.

Developed a continual learning platform with FastAPI and React, facilitating real-time model training and updates from streaming data.

Unified frontend applications and shared components within Turborepo, increasing code reuse by 40% and decreasing build time by 30%.

Extended Prometheus's Go parser to include created-timestamp support, enabling accurate rate calculations for previously unsupported metric streams.

Implemented in-memory parser caching and targeted microbenchmarks, enhancing memory throughput by 300%.