What trAIce doesn't do
We're a cost & margin product for production AI. The problems below are real and worth solving — they're just not ourproblems. Here's who solves them well.
Being explicit saves us both time: if you came here for something on this list, one of the tools we link is a better fit.
In-session token compression
Reducing token waste during a long Claude Code or Cursor session — bloated CLAUDE.md, verbose command output, late compaction.
Different audience (devs using AI to write code) and different mechanism (in-session compression). We're cost+margin for production AI traffic.
AI gateway / routing / fallback
Sitting in the request path to route between models, retry on failure, enforce auth, and abstract providers.
Gateway is a crowded category and a different commitment. Our explicit non-goal is "we do not sit in your request path."
Evals platform / prompt versioning / playground
Defining datasets, comparing prompt/model variants side-by-side, scoring with autoraters, managing a prompts library.
We adopt the narrowest possible eval primitives only in service of cost-validated savings (Phase G). We don't build a workbench.
Span-level observability / traces
Distributed tracing of LLM calls, latency histograms, error budgets — full APM for AI apps.
Our wedge is cost attribution and customer margin, not span-level observability. We complement; we don't replace.
Prompt management / versioning
Storing prompts as versioned artifacts with branching, rollback, A/B tests.
Prompts belong in your code repository under the same review process as the rest of your application.
AI gateway billing / metering
Charging your own customers for AI usage, generating per-seat invoices, mediating provider quotas.
We measure cost and margin; we don't run the billing pipeline.