How SysGeeks uses AI
Most of SysGeeks works without an account or AI. AI is an optional feature for signed-in users. This page says what it does, what is sent, which models run, how long data is kept, how its output is labelled, and when credits come back.
Where AI is used
Free practice does not use AI. Choices are checked against an authored key, written answers can be self-assessed, and the model check on a project is deterministic code. None of these need an account.
Signed-in users can also run the features below. Each one shows its credit cost, and your balance after it, before it runs. There are no subscriptions, no automatic top-ups and no recurring charges.
| Feature | What it does | Cost |
|---|---|---|
| Assessment of a written answer | Assesses your written answer against the rubric points of its question. Every assessment quotes the sentence it relied on, and the server rejects a quote it cannot find in your answer. | No credits, up to 20 a day |
| Draft from a description or README | Drafts a project’s architecture from text you give it. | 1 credit |
| Draft from a public GitHub repository | Drafts a project’s architecture from the files you choose in a public repository. | 2 credits |
| Architecture review | Reviews a project’s architecture and reports what it finds. | 2 credits |
| Tailored question set | Writes 8 to 12 questions about your own design, covering trade-offs that the standard questions cannot know about. | 1 credit |
Not available yet: an AI interviewer that asks follow-up questions in a mock interview. It is planned for a later phase.
What is sent
- Your text. The answer being assessed, or the description or README you give for a draft, is sent when you run that feature.
- Repository files. You see the files before the run starts, and only the files you select are sent.
.envfiles (other than.env.example) and*.pemand*.keyfiles are never fetched. Values that look like secrets, such as AWS keys, GitHub and Slack tokens, private key blocks andpassword=-style assignments, are replaced with[REDACTED]first. Matching works by pattern, so it can miss a secret. Do not select files you would not share. - An account identifier. Each request carries an internal account ID, which is used to attribute its cost.
Which models run
AI calls go through Vercel’s AI Gateway. Each feature has a primary model and a fallback from another provider, which runs if the primary fails.
- Drafts, architecture reviews and question sets: Claude Sonnet 5.5, with GPT-6.1 Sol as the fallback.
- Assessments of written answers: Claude Haiku 5.5, with GPT-6 Luna as the fallback.
Each AI output records the model that produced it. Models can change, and the model named on a report is the one that ran.
How long data is kept
- Repository file contents are not stored. Only the list of file paths and the output are kept.
- Each run keeps its output, the model, token counts, estimated cost and any error code. AI reports are stored under your account, so you can open them on another device.
- Your learner state, meaning your answers, progress, projects and tailored questions, stays in your browser. Syncing it to our servers is not available yet.
- Providers are asked not to train on prompts. By default, requests are routed only to providers that keep no data (zero data retention).
- You can delete your reports and run data at any time from your account.
How AI output is labelled
Every piece of AI output says what the AI did. These are the labels you will see:
- AI draft a draft the model wrote from your input.
- AI review a review of your design.
- AI-assessed an assessment of a written answer.
An AI assessment shows its basis as ai-assessed, next to the result. The other bases are checked (against an authored key), self-assessed (you marked the points) and mixed. You can always mark the points yourself instead.
Each report carries a line like this:
AI-generated by {model} on {date} from {inputs}. It can be wrong; low-confidence items are guesses.
Draft items are marked stated, inferred or guessed, so you can tell what your text says from what the model filled in. Once you accept a draft, it is marked as AI-drafted and edited by you. That mark is not removed silently.
Credits and limits
- Welcome credits: 5, granted once after your email address is verified.
- Packs are one-time purchases: 15 credits for $5, 40 credits for $12, 100 credits for $25. Credits never expire. The home page lists them in full.
- Assessments are limited to 20 a day, or 100 a day after any purchase. Drafts, reviews and question sets are limited to 10 runs a day each.
- There is a daily cap on estimated model spend. Passing it pauses free assessments, and passing three times it pauses all AI. The cap resets each day.
When a run fails or is wrong
- Failed runs. If the provider errors, the run times out, or the output is still invalid after one retry, the credits come back in the same response, and the page says so.
- Runs that stop unexpectedly. If a run is cut off before it finishes, its credits come back the next time your account loads, once the run is 10 minutes old.
- “This is wrong about my system.” Every report and question set has this button. It refunds that run’s credits straight away. Each run can be flagged once, and an account can flag up to 3 runs in any 30 days. Beyond that, email support.
- Purchases. A pack with no credits used can be refunded within 14 days of purchase. The refunds policy has the details.
Questions about your data
The privacy policy covers who processes data and how to ask for it to be deleted.