Notes from
the shop floor and
the production stack.
Deep engineering write-ups on building agentic infrastructure for manufacturing: voice agents, semantic layers, model economics, and the tradeoffs nobody markets honestly. Every number cited.
The Manufacturing Context Layer
An architecture for governed, agent-ready operational context.
The Manufacturing Context Layer
Manufacturers have spent a decade centralizing plant data — but the bottleneck to production AI isn't connectivity or storage. It's context. A reference architecture for governed, agent-ready operational context: connectors, an ISA-95 knowledge graph, and governed agents.
The Voice Agent Tradeoff Triangle
A production-engineering view, and a CTO procurement view, of what it actually takes to ship a voice agent in 2026. The four-way tradeoff between cost, latency, language, and intelligence, in three parts. Every number cited.
- 1The Pipeline & the Language WallWhy a production voice agent is a four-way tradeoff nobody markets honestly. How the pipeline fits together, where the milliseconds go, and the language wall that hits the moment you leave English.6 min read§1–§2 · pipeline + STT
- 2The Inner Triangle: LLM & TTSThe two stages where cost, latency, and intelligence pull hardest in different directions. Picking the right LLM, why prompt caching changes everything, and the dirtiest fight in voice AI: text-to-speech.6 min read§3–§4 · LLM + TTS
- 3The Tradeoff, the Bill & the PlaybookIt all collapses into one picture. The four-axis radar, what a real deployment costs at 8,000 voice-minutes a month, the latency and cost levers that actually move the needle, and the honest conclusion.8 min read§5–§9 · tradeoff + cost
Building voice agents for the floor?
If you want a deep-engineering tour of a production voice agent running in Indian manufacturing, from work orders to preventive maintenance to WhatsApp escalation, talk to us.