Baseline
Review available usage, token or call records, latency traces and billing for the selected workload.
Agentic applications · 01 · The problem and the decision
Model routing and inference cost reviewWhat does one conversation cost you, to the cent?
The inference bill is rising, but cost per useful task and latency are hard to explain. Establish a baseline and decide whether routing, caching or another change is worth testing.
For CTOs, AI platform owners and the people accountable for an application's inference spend.
1–2 weeks elapsedOne agreed AI workload and measurement scopeA named architect accountable
02 · What we check
Agree the scope. Inspect the evidence.
Agree the workload sample, available evidence and measurement depth. Review cost alongside task quality and response time, with the limits of each comparison stated.
Routing and caching
Examine task classes, repeated context, cache suitability and fallback. Identify candidates to test.
Tradeoffs
Compare the agreed options against quality, latency, data handling and cost; distinguish measured results from forecasts.
What we need from you
- Usage and billing data, representative requests and available traces.
- An application owner who can define useful output and acceptable latency.
- Agreed providers/models, data restrictions and a budget for any permitted measurement runs.
Scope and boundaries
- A broad model leaderboard, gateway implementation, detailed implementation-ticket engineering and a guaranteed saving are excluded.
- Additional workloads, providers, measurement runs or remediation require separately agreed scope and cost.
03 · What you get
A decision you can act on.
- A cost-per-task and latency baseline with workload and method stated.
- A routing/caching options record showing quality and operational tradeoffs.
- A recommendation with unknowns and the evidence needed before implementation.
- A measurement plan for checking the effects of any subsequent change.
04 · Relevant work and a sample
See the thinking behind the review.
An application context for the engineering questions
Atsky Voice is an Atsky demonstration developed in partnership with Telnyx. It provides context for questions about inference, latency and tool use, not a promise of the same performance or cost in a client workload.
05 · Duration and price
A defined engagement. An agreed scope.
From €5,500Starting price, excluding VAT
From €5,500, excluding VAT. We quote a fixed fee after agreeing the workload and model inventory, data sample, access and measurement depth. The proposal confirms the elapsed delivery window and any run costs. Extra measurement, remediation, implementation and detailed ticket engineering are separate.
We agree the included scope, deliverables and fixed fee in the proposal before work starts.
- Duration
- 1–2 weeks elapsed
- Base scope
- One agreed AI workload and measurement scope
- Your data
- Read-only access, in a workspace that holds only your engagement. If your policy requires it, we work with EU-only processing and zero data retention. Any extra cost is named in the proposal.
Our own AI-assisted tooling collects evidence and drafts findings; a named architect checks and signs them. How we work and handle data →
06 · Request a working session
Start with the question you need to resolve.
In 45 minutes, discuss the workload, the spend question and the evidence available. Establish which measurements would change your decision.
This requests a 45-minute conversation. We confirm a suitable time with you.
Are broader architecture and control questions blocking the pilot? Explore the agentic architecture review.