Recommendations
Engineering cost recommendations with evidence and math
A dashboard shows how much was spent. Nythrex recommendations tell you where money produced no result and what to do about it. Every recommendation is the result of a named rule over your records: with evidence, arithmetic and a check of whether the fix worked.
AIAlready spent15 sessions
AI sessions with no commit$37.00
5 of 15 finished AI sessions produced no commit — half of the AI spend for the period.
How it was calculated
- A session counts as finished after a few hours without activity.
- Small sessions below the threshold are not counted.
- This is an upper bound: research and reading code land here too.
What you can do. Check whether long sessions end in a commit and whether the hook that signs commits with the session is installed.
The rules are real; the examples are demo data.
Rules that already work
Recommendations are grouped by area: AI, CI, cloud and process. Below are the rules from the example above, with what each of them counts.
| Rule | Area | What it counts |
|---|---|---|
| AI sessions with no commit | AI | A session counts as finished after a few hours without activity. |
| Expensive model on helper requests | AI | The same tokens re-priced at the next model tier down. |
| Context written to cache again | AI | A rewrite is a large cache write after a pause longer than the cache lifetime. |
| Several workflows per push | CI | The most expensive workflow of each commit is treated as necessary. |
| Failed and repeated runs | CI | Failed means a failure or timed_out conclusion. |
| Cloud spend with no owner | Cloud | The total across subscriptions and resource groups not assigned to any project. |
| Abandoned work | Process | A unit is abandoned if its pull request was closed without a merge or has had no activity for a long time. |
How every recommendation is built
Records
Links to the specific sessions, runs or resource groups the conclusion is based on.Arithmetic
How the amount was obtained: which records were counted, which were left out and why it is an upper bound, if it is one.Basis
“Already spent” or “estimate” — so that the potential is never confused with the actual.Next step
What exactly can change: a setting, a process, or a check of whether the resources are still needed.
What recommendations do not do
- they do not promise savings — a potential amount is never added to what was saved;
- they do not rate people or look for someone to blame — the rules look at process and settings;
- they do not invent conclusions from too small a sample;
- they need nothing to be run — they fire on their own after every data refresh.
Perguntas frequentes
Are the recommendations generated by a language model?
No. Every recommendation is the result of a deterministic rule: the same records always give the same recommendations. Plain-language wording is optional, and the model receives only the facts of one recommendation — no records and no code.
What if there is too little data?
The rule stays silent and says how many more records it needs. Nythrex does not draw a confident conclusion from three data points.
How do I check that a recommendation worked?
Mark it as fixed. After at least a week a before-and-after comparison on the same records appears.
Can I get recommendations by email?
Yes, a weekly digest — optional, with one-click unsubscribe in every email.
Leia também
Deixe de adivinhar.
Comece a ver.
Sem cartão e sem subscrição: os planos pagos ainda não foram lançados.