Lokesh R_

An AI lead qualification and follow-up agent for high-ticket sales — one tool-calling agent that qualifies inbound leads, scores them, and routes them to a rep who can take over the chat mid-conversation.

• A single agent loop written against the LLM's native tool-use API — no LangChain — that normalises free-text chat into structured filters and calls fixed tool signatures.
• Scoring and assignment are deterministic Python, never LLM calls: rules-based Hot/Warm/Cold tiers and tier-weighted round-robin routing, so both stay auditable.
• Live human takeover over Server-Sent Events — the agent goes silent the moment a rep steps in, on either a matched intent or an employee reply.
• Celery beat stall checker drafts a follow-up email for leads quiet past 48h and pings the assigned rep to approve, edit, or send.
• Parameterised indexed Postgres for property search, with pgvector reserved for the separate document knowledge-base RAG.

Reading about DynamoDB's internals got me curious enough to build my own — a masterless distributed key-value store in Go, working through the Dynamo paper one section at a time.

• Consistent-hash ring with ~150 virtual nodes per physical node, so resizing the cluster moves one arc of keys instead of remapping everything. 500 keys across 3 nodes land 183/162/155 — within ~9% of even.
• Every node coordinates every request against the key's preference list — no primary to forward to, so a down replica never makes a key unwritable.
• Quorum reads and writes over N replicas, with R + W > N enforced at config load so a config that can't guarantee overlap is rejected on startup.
• Hinted handoff and sloppy quorum keep writes available during a failure: an unreachable replica's copy is parked on a live stand-in and handed back automatically on recovery.
• Vector clocks on every value — the coordinator bumps its own entry per write, replicas compare before overwriting, and clocks survive WAL replay after a crash.

A concurrent WebSocket server in Go built for backpressure resilience.

• Dedicated read/write goroutines per connection, preventing slow clients from blocking server resources.
• Weighted priority job queue (3:1 high-to-low) so heartbeat and ping messages are never starved by application payloads.
• Non-blocking outbound queuing with a 100-message buffer per client, dropping on overflow to stay responsive under backpressure.
• Graceful shutdown on SIGINT/SIGTERM — broadcasting warnings, draining in-flight jobs, and safely closing the worker pool.

A full-stack travel blogging platform with a decoupled client-server architecture.

• Google OAuth authentication with JWT-based session management.
• Global application state managed with Redux Toolkit across authenticated user flows.