The free AI-spend assessment — what to expect

A working session, not a sales call. We measure NeuralMind’s actual token-reduction ratio on one of your repos and hand you a spend model built from your numbers — whether or not you deploy anything.

Book it: hello@neuralmind.uk

Fastest path: run it yourself in 5 minutes

You don’t need to wait for a call. The assessment’s core measurement is a command you can run right now, on your own hardware, with nothing leaving your machine:

pip install neuralmind
neuralmind build .          # index your repo locally
neuralmind benchmark .      # → your measured reduction ratio + cost estimate

Email the benchmark output to hello@neuralmind.uk along with three numbers — engineer count, monthly API spend, and GPU fleet if you run one — and we’ll send back the full three-line spend model in your numbers. You share only the report; your code stays where it is.

What you get

1. A measured reduction ratio on your code. We run neuralmind benchmark against a repo you choose (on your hardware if you prefer — nothing needs to leave your machines). The output is per-query input-token counts vs. a naive baseline, the same methodology as our public CI benchmark.
2. A three-line spend model in your numbers. Per-seat subscriptions, usage-based API spend (direct, OpenRouter, Bedrock, Vertex), and self-hosted GPU capacity — modeled separately, with the assumptions written down so your finance team can change them. Structure documented in BUSINESS-CASE.md.
3. An honest fit verdict. If your workload is generation-heavy, your repos are small, or prompt caching already covers you, the model will say so. The methodology’s failure modes are public: The case is weaker if…

What we’ll ask — have this ready

The model is only as good as its inputs. The call goes fastest when you bring:

Team & tools

• How many engineers use AI coding tools, and which ones
• Seat prices and whether your plan has usage tiers or overage charges

API / usage-based spend

• Last 3 months of invoices from each provider (OpenRouter, Anthropic, OpenAI, Bedrock, Vertex — whatever you use)
• Input-vs-output token split if your dashboard shows it
• Rough share of traffic that’s code questions / agent retrieval vs. other workloads
• Whether prompt caching is enabled

Self-hosted inference (if applicable)

• GPU types and count (e.g. H100s), and your contracted $/GPU-hour or amortized cost
• Fleet utilization, and prefill-vs-decode share if you’ve profiled it
• Queries/day served and typical context length

Workload shape

• Size of the repo(s) you’d point NeuralMind at
• Code questions per engineer per day, roughly

Don’t have all of it? Come anyway — we’ll model with stated assumptions and mark every estimated input as estimated.

What we won’t do

• Upload your code anywhere. The benchmark runs locally; you can run it yourself and just share the output.
• Quote savings we didn’t measure. Measured numbers are labeled measured; derived numbers are labeled derived, same as everywhere else in this project.

The software is MIT-licensed and free. Commercial support for deployment, integration, and evaluation is available — ask on the call.