Vinay Kumar

Vinay Kumar

Founder & CTO

About

Your AI bill has a dirty secret: you're paying for tokens that make your answers worse. Research shows accuracy can drop 15–20 points when key facts get buried mid-context (lost in the middle), agents accumulate bloated tool outputs turn after turn, and Anthropic reports multi-agent systems consuming ~15× the tokens of ordinary chat. Cost, latency, and quality all degrade together — a phenomenon now called context rot.

This is a hands-on session on context engineering for the token economy: filling the context window with just the right information and nothing else. Problem first, tool second — we start from a real token bill, install a token meter, then apply the four levers used by production agent teams: Write (scratchpads, memory files), Select (RAG, reranking, tool retrieval), Compress (summarization, compaction, output stripping), and Isolate (subagents, sandboxes, state). We'll restructure a prompt for KV-cache hits to slash input costs on hot paths, and demo the token tax nobody budgets for: Indic-language text costing several times the tokens of English, and image inputs billed by resolution.