Roshan Singha

About Roshan Singha

Frontend engineer specializing in real-time, data-intensive UIs — streaming AI interfaces, WebSocket-driven dashboards, and complex state management at scale. Currently building the frontend and AI integration layer for an enterprise digital twin platform (SSE streaming across OpenAI/Claude/Gemini, RAG-powered chat, credit metering). Previously: real-time kitchen operations system at Rebel Foods serving 400+ kitchens.

Experience

1. Sde-2

   Primathon

   Jun 2023 — Present · Gurugram, IN

   AI & RAG→ Designed RAG pipeline end-to-end — chunking, vectorization (40-concurrency queue), embedding storage; prompt-resolver augments LLM calls with retrieved context across OpenAI, Claude, and Gemini→ Multi-provider LLM orchestration with SSE token streaming, provider-specific prompt formatting, and fallback routing on provider failures→ Credit-gated AI metering — atomic pre-deduction with auto-refund on failure, handling race conditions across async provider callsFrontend→ Built the real-time AI chat interface with SSE streaming, inline data visualization, JWT refresh mutex, cross-tab session sync, and optimistic updates→ Voice onboarding via SSE — 8+ phase state machine, event deduplication, recoverable errors, MediaRecorder capture with WebM→WAV conversion→ Chat orchestration hook managing SDK lifecycle, WebSocket connections, JWT refresh, multi-session history, media streaming, citation extraction, and tool call routing→ Dataset upload hook with exponential backoff reconnection, JWT expiry-aware refresh, batch completion detection — zero dropped uploads in production→ WCAG 2.1 accessible modals with inert-based focus trapping, aria-live regions for streaming updates, and full keyboard navigation across all flows→ Responsive dual-rendering via PlatformRouter — mobile/desktop component trees across 30+ routes with code-split bundlesBackend→ SSE + WebSocket hybrid real-time layer scaled via Redis pub/sub across multiple container instances→ SSE proxy bridging Python AI service to frontend — multi-subscriber fanout, event buffering with replay, heartbeat keepalive, and Redis-backed session recovery→ 10 BullMQ processors for vector uploads (40 concurrency), conversation storage (50), KG extraction, and voice training — 92-table PostgreSQL schemaHard Tradeoff→ Fixed critical SSE bug where React StrictMode double-unmounts killed active streams — lifecycle refs distinguishing cleanup from teardown, eliminating false terminations

Education

• Kendriya Vidyalaya

  Higher Secondary School Certificate, Science

• National Institute of Technology Agartala

  Bachelor of Technology - BTech, Computer Science

  2015 — 2019