Prismatic Labs

Prismatic Labs · Inference Waste Control

Production AI silently wastes inference.

Most of it is invisible: an agent stuck looping, retrieval dragging in context it never uses, a cache quietly missing. The bill is usually the first to notice.

And it adds up. Every wasted call still costs money and still draws power, carbon, and water from a data center somewhere, and the bigger your AI gets, the bigger the pile.

Vetch reads your own metadata, finds the waste, and stops the loops that are running for nothing.

No generic savings claim. Vetch measures your traffic, then estimates avoidable waste from your own metadata.

Mini clicker game: simulated waste only

Vetch · waste clicker 0 calls
Waste cards pop up. Click Stop waste to build your streak. Demo data only. Not a benchmark or savings claim.
Waiting for the next waste card…
Score0
Demo waste blocked$0.00
Cleared0
Streak0×
Simulated demo. Not a guarantee of savings.

Observability shows what happened.
Vetch can stop confirmed stalled loops.

Programs & affiliations

Programs we’ve taken part in.

Selective programs for technical AI startups. Participation doesn’t imply product endorsement.

NVIDIA Inception program member

Member of NVIDIA Inception

A startup membership program for companies building with AI and advanced computing.

Incubator

Took part in the Oxford University Innovation Incubator programme

Cohort 15, Phase 1. Ran from 26 March to 1 April 2026.

Open source · Apache 2.0 · Python · pip install vetch

One import. Every inference call tracked, attributed, and ready to stop.

01 · Detect

Waste patterns in metadata

Stalled loops, RAG bloat, cache misses, unattributed spend: all derived from call metadata. No prompt text stored.

02 · Attribute

Cost by feature, customer, team

Tag every call. Cost, energy, carbon, and water accumulate per session, per tag, per workflow. “The bill went up” becomes “the RAG search feature for enterprise customers did it.”

03 · Control

Warn · Kill · Reroute

Observe first. Promote high-confidence advisories to production controls. Fail-open always. If Vetch fails, your inference continues.

Choose your next step

Run Vetch yourself, or bring us in when the evidence needs a decision.

Install the open-source SDK for free. If a scan surfaces patterns you want help interpreting, Prismatic Labs can review the evidence, design attribution tags, and plan safe production controls.

Run it yourself · Free

For engineers who just want to see what’s burning. Install the SDK, tag your calls, and watch the waste advisories come in. Warn-only, Apache 2.0, yours to keep.

Run free scan →

Review & Audit · from £295

For teams that want the evidence turned into a decision. We read your Vetch output and hand back a prioritised, spend-saving action plan. A snapshot from £295, a full audit from £950.

Request a review →

Production controls · from £2.5K/mo

For teams ready to stop the waste, not just watch it. Kill, reroute, and budget guardrails, with rollout planning, monitoring, and dashboards. Carbon and CSRD-ready reporting come off the same data.

Talk to us →

Starting prices. Larger or regulated engagements quoted after discovery.

Prismatic Labs

Inference waste control for production AI.

marco@prismaticlabs.ai

© 2026 Prismatic Labs

© 2025 NVIDIA, the NVIDIA logo, and NVIDIA Inception are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries.