Arjun Gupta

Stealth AI Startup

• •

  Owned three production AI codebases end-to-end as sole full-stack engineer and product owner, managing async backends and frontend deployments (Render, Vercel), enabling the team to ship the v0 pilot to live users without an additional engineering hire.

• •

  Architected a federated multi-tenant agentic platform (Hub/Spoke model) with A2A communication and stateful LangGraph orchestrators backed by a custom 14-tool MCP server, enabling spoke users to query a hub's relationship graph through agent-mediated consent without exposing private signals.

• •

  Engineered an async-first Python stack (FastAPI, asyncpg) with an event-driven message bus and deterministic dedup, handling real-time WhatsApp WebSocket ingest and Meta Cloud API delivery at sub-second consumer lag with at-least-once guarantees.

• •

  Designed a 4-way hybrid retrieval engine fusing PostgreSQL pgvector (HNSW + FTS) with a perception-aware Neo4j knowledge graph and HyDE + Cohere reranking, improving top-3 candidate recall over baseline vector-only search on internal eval queries.

• •

  Built stateful GPT-4o workflows with strict-JSON tool use and an LLM-as-judge (PageIndex tree) for intent classification and hallucination control; enforced multi-tenant data privacy via Clerk JWT auth and workspace-scoped SQL filters to prevent cross-tenant leakage by design.

ZuduAI

• •

  Shipped production AI agents on LangChain + LangGraph in a microservices setup, delivering RAG workflows that served live customer queries across enterprise deployments.

• •

  Designed and deployed low-latency voice AI systems on LiveKit + PipeCat integrating Deepgram STT and multilingual TTS, achieving real-time response latency suitable for natural conversational interaction.

• •

  Built post-call NLP pipelines (sentiment, entity extraction, intent) on Azure Cognitive Services orchestrated via Apache Airflow + n8n, generating structured insights from raw call transcripts to feed downstream analytics.