Personal Blog — Suman Jana

Handling 50+ Tools in AI Agents Without Burning Your Token Budget

50 tools kill your token budget. Dynamic selection, schema compression, and routing for production AI agents.

GraphRAG vs LightRAG — community detection, knowledge graphs, and structured retrieval beyond naive vector search.

Stop vibe-checking your LLM app. RAGAS, DeepEval, and PromptFoo for evals that catch regressions before production does.

WebSockets, PCM audio, VAD, and tool calling — building real-time voice AI with the Gemini Live API from scratch.

ColPali encodes document images directly — no OCR, no chunking. Here's how to build multimodal RAG that actually works.

LangGraph, CrewAI, AutoGen compared — stateful agent workflows, circuit breakers, and patterns that survive production.

Winning 100agentdev over 600 teams — fan-out architecture, a 3 AM loop bug, and lessons from 48 hours of hacking.

Safely deleting 90K PII records from live Postgres with zero downtime, batched deletes, and a complete audit trail.

IntelliSupply — LLM agents, graph simulation, and structured handoffs for supply chain disruption prediction.

How I built a multi-tenant AI chatbot platform using Vercel AI SDK and Gemini at a hackathon.

How I am planning to reach full financial independence as a full-stack developer.