Agentic and voice AI systems — multi-agent orchestration, persistent graph memory, acoustic emotion detection, and real-time CRM integration.
Voice agent on LiveKit Agents 1.6 that remembers callers, interprets emotional tone via acoustic analysis, and adapts response style dynamically. Trust-weighted Neo4j knowledge graph with contradiction decay and sarcasm detection.
For each prospect it mines public pain signals, generates a personalized microsite + working prototype deployed to Cloudflare Pages, records an ElevenLabs-narrated Loom-style walkthrough, and drafts the outreach into a Convex-backed human-review console. Built on Hermes (Nous Research) for the GrowthX × Hermes buildathon.
45-second client briefings via phone call with portfolio health, risk flags, and cross-sell recommendations. Live whisper coaching during calls, voice command CRM execution, and post-call analytics across 7 CRM integrations.
Production-grade backend systems built from first principles — distributed architectures, geospatial indexing, and fault-tolerant services. Engineering write-ups available on Medium.
Real-time geospatial driver discovery using Redis GEOADD/GEORADIUS with sub-200ms P95 driver lookup. Event-driven assignment flow with Pub/Sub designed for horizontal scalability under high concurrent-ride load.
Event-driven file synchronization across nodes with chunk-level diffs to minimize bytes-on-wire. Conflict-resolution guarantees with zero data loss under concurrent writes across distributed clients.
From-scratch L7 load balancer with round-robin + weighted routing, health-check probing, automatic failover, and a pluggable strategy interface. Backed by Redis Pub/Sub for live node updates.
Classified financial emails using GPT-4 few-shot learning on a 180-email dataset. Reduced redundant LLM calls 40% via cosine-similarity caching (threshold >0.95) in MongoDB Atlas Vector Search. Scaled with FastAPI async + Kubernetes HPA.
Deep dives into distributed systems, system design patterns, and backend engineering — published on Medium.
A deep dive into a real-time, scalable architecture with FastAPI, Redis, and event-driven design.
What actually makes distributed file synchronization hard — the core engineering challenges behind eventual consistency.
A step-by-step engineering walkthrough with Go, Docker, Redis, and Pub/Sub.
Practical guide to the Factory Method pattern using Python for backend development.
Starting with Singleton — introduction to design patterns for backend engineers preparing for system design interviews.
I'm Himanshu, an AI & backend engineer based in Bengaluru, India, shipping LLM features in production fintech at Wells Fargo. I build AI Teammate, a Python LLM + RAG chat assistant that lets 15,000+ financial advisors query internal systems in natural language and act on their CRM through MCP tool calls.
Day to day that means Python, FastAPI, LangChain, and RAG pipelines on event-driven, cloud-native infrastructure. Outside work I build voice agents — a cognitive voice AI with persistent Neo4j graph memory and real-time acoustic sarcasm detection, and SYNC, a live call co-pilot that coaches relationship managers mid-conversation.
I write engineering deep dives on Medium — ride-matching at Uber scale, Dropbox-style file sync, and load balancer internals. Outside of that: chess, the gym, and shipping side projects. Always open to conversations about AI systems and remote-friendly opportunities.
