Om
Avchar

Designed a two-tier traffic routing system using Priority Queues and a custom Load Balancer to separate free-tier and premium API calls — required by the constraint of a single shared OpenAI rate limit across 550,000+ users.

Implemented a Redis-backed caching layer ('Emotional Cache') that stores per-user conversation context as structured embeddings, reducing redundant LLM inference calls and cutting average response latency by ~40% under peak load (1,000–3,000 DAU).

Architected the message queue to decouple real-time chat from background memory writes, enabling the system to sustain throughput without blocking on database I/O.

Built a multi-modal pipeline that transcribes, chunks, and summarizes YouTube video content into structured PDF notes in 7–8 seconds for an 8-hour video — achieved by parallelizing transcript segmentation with async worker threads rather than sequential processing.

Implemented a 'Deep Note' generation mode that runs a secondary Gemini API pass to produce visual aids and diagrams (10–16 sec total); the added pass is only triggered on explicit user request to keep base latency low.

Integrated Google Drive API for cloud storage and delivery, with a custom PDF rendering engine that formats structured output into student-ready notes without a third-party PDF library.

Built a scheduled content pipeline using cron-based workers that poll Instagram and X trend APIs every 15 minutes, generate brand-specific post copy via LLM, and queue posts for automated publishing — processing ~96 trend checks per day with no manual triggers.

Designed the backend to be stateless across workers, with MongoDB storing trend history and generated content; this allowed horizontal scaling without coordination overhead.