Debabrata Acharjee is Head of AI Innovation at Matific, where he leads production LLM systems and the company's agentic architecture strategy for a K-9 math platform used by more than five million students worldwide. Before that, he was Co-Founder and CTO of EventBookings, a global event-management SaaS platform, where he built and led the engineering organization through a full platform migration and a period of substantial revenue growth.
Earlier still, he spent eleven years at 3P Learning helping build and scale Mathletics, one of Australia's most significant edtech platforms — from early startup through ASX listing, reaching more than five million students across 159 countries.
His path to AI infrastructure runs through 26 years in technology overall — banking, insurance, automotive, and edtech. The Six Laws and the optimization techniques in Production LLM Architecture are derived from architectures he has shipped, scaled, and debugged in that time — not from the literature.
This book distills two years of production LLM engineering specifically: the cost overruns and latency spikes diagnosed under load, the cache invalidations debugged at 3 a.m., and the framework that emerged when six recurring failure patterns turned out to be the same six constraints applied differently. He builds teams the way he builds products — hire for curiosity, create space to experiment, hold the bar high.
Every production LLM system I built failed the same ways.
Latency spiked because context grew without bounds. Costs exploded because every query sent the entire knowledge base through prefill. Quality degraded because the system prompt accumulated patches until nobody could read it. Cache hit rates collapsed because a timestamp appeared before the static prefix.
These were not model problems. They were engineering problems — and the literature had nothing for them. The tutorials covered prompting. The papers covered architecture. The blog posts covered the demo. Nothing covered what happens when the system takes real traffic.
I traced every failure I encountered to one of six root constraints — constraints imposed by transformer mechanics, hardware economics, and information theory. A latency spike, a cost overrun, and a quality regression looked like three separate incidents until I mapped each to a violated constraint. Six constraints explained nearly every production issue. They became the Six Laws.
This book is the production reference I could not find during two years of building, breaking, and fixing LLM pipelines. The cost tables reflect real bills. The failure modes are ones I diagnosed on live systems. The optimization stack compounds the way I describe because I measured it compounding.
The field moves fast. Prices drop. Windows expand. Providers launch quarterly. I built this book on structural constraints — prefill mechanics, token economics, caching invariants — that remain stable as the surface details shift. The specific dollar figures will date. The Six Laws will not.
— Deb
Fourteen chapters, six governing laws, and the production reference that came out of shipping this work.