Insights & Architecture
Field notes on engineering, agentic AI, and digital strategy from the VigyanBytes team.
A 35B model at 237 tokens/sec in 13.6GB — here's the stack
Expert pruning, multi-token prediction, and quantised KV cache stacked together get a 35B mixture-of-experts model to 237 tok/s in 13.6GB on a single consumer RTX 5090.
AI tools aren't rivals — they're job descriptions
Running several AI coding tools side by side on real infra work and giving up on ranking them — one thinks, one suggests, one executes. They're roles, not rivals.
I built in private too long and called it discipline
Months of building in private, telling myself one more governance rule needed cleaning up first — until the build log itself became part of the product, not a delay.
I didn't leave the cloud for the bill
Running your own AI infrastructure isn't about the bill — it's the operational knowledge you never build when every failure is absorbed by someone else's abstraction.
I rejected framework-shaped comfort — here's the orchestration I trust at 2am
Why plain-text governance, explicit handoffs and phase gates beat an orchestration framework when infrastructure breaks at 2am and you need to know exactly what happened.
"Supported" is the most misleading word in AI infrastructure
Twenty-four hours trying to run a 31B model on a consumer 32GB GPU through vLLM — every fix applied, still 112MB short. Supported isn't the same as it actually runs.
I wrote storage law before I wrote storage code
Why a private AI cloud needs storage rules before storage code — mountpoints as a contract, closed storage classes, and drift that gets reconciled, not absorbed.
Three AI tools, one infrastructure, no orchestration platform — just plain-text governance
Running Claude, Gemini and Codex against the same infrastructure without an orchestration framework — just a shared, plain-text constitution they read before acting.
Prompt caching in production: exact vs semantic, and the break-even math nobody does
Exact vs semantic prompt caching, the break-even math that decides whether caching saves money or costs more, and why a zero on a local vLLM cache-hit field can be lying.
I built a private AI cloud from an old desktop and a new GPU — here's what actually broke
Running an RTX 5090 next to a seven-year-old desktop as a two-node private AI cloud — the flaky router port, the Bluetooth dongle bug, and what actually broke first.
Stop running AI pilots. Start shipping production agents.
We help enterprises and founders who need architecture thinking, not just prompt engineering.