Full Stack and AI Systems Engineer with production experience building RAG pipelines, agentic AI systems.

leetcode

github

Manish Kumar  • Peerlist

Bastion is a multi-agent code intelligence platform that scans GitHub repositories for security vulnerabilities and surfaces actionable fix recommendations. Architected a 6-stage async scanning pipeline using Celery and Redis (clone → validate → Bandit security scan → AST parse → AI review → save findings), enabling parallel multi-file analysis without blocking API requests. Built real-time scan progress updates via Redis Pub/Sub and WebSockets. Integrated a RAG system using Pinecone with AST-based symbol-level chunking for repository-aware code understanding. Combines Bandit static analysis with Gemini AI review behind GitHub OAuth and JWT-protected endpoints.

Bastion

Chronel is a production RAG (Retrieval-Augmented Generation) platform that enables semantic querying over uploaded PDFs. Built a distributed 3-service architecture with independently deployed frontend, backend, and AI inference services on Render, enabling zero-downtime per-service updates and horizontally scalable async inter-service communication. Engineered the retrieval pipeline with chunk size 300 tokens, Pinecone top-k=5 vector search, and cross-encoder reranking to top-3 chunks before LLM context injection - improving answer relevance and minimising hallucinations. Measured end-to-end latency of 1.5-2s via LangFuse: 60% in vector retrieval, 40% in LLM generation.

Chronel

CinemaGhar is a full-stack movie ticket booking platform built for reliability under load. Reduced API latency by 40% (300ms → 180ms) using Redis cache-aside with TTL-based invalidation. Validated concurrency safety under 100 simultaneous booking requests via k6 — exactly 1 booking confirmed, 99 rejected with 409 conflict, and 0 server errors, enforced via Prisma transactions with atomic seat conflict detection. Features role-based access for admins, cinema managers, and users, seat lifecycle management with cron-based release of abandoned bookings, and mocked Stripe payments.