Who it is for
A SaaS that shipped an AI feature quickly: no idea what each feature costs, no hard limit on spend, and customer data going straight into prompts.
The problem
- AI bills grow faster than usage: every call re-sends instructions, history and documents, and retries are billed again.
- Personal data in prompts and logs worries customers, security reviews and privacy laws.
- When the model fails, is throttled or returns broken output, the feature breaks with it.
What I do
- Token and cost logging per feature, and provider budgets with alerts.
- Redaction of personal data and secrets in code, before the call, with tests.
- Provider guardrails (Bedrock Guardrails or moderation) scoped to user input, and strict validation of the model's output.
- Hard daily and monthly caps checked before each call, with a safe fallback when anything fails.
How it works
- 1Map every AI call
- 2Log tokens and cost per feature
- 3Redact personal data before the call
- 4Guardrails and strict output checks
- 5Hard daily and monthly caps
- 6Safe fallback on any failure
Day by day
- Day 1Map every AI call; add cost logging per feature; compare your real usage with a cost model.
- Day 2Hard caps in code with a fallback; budgets and alerts.
- Day 3Redaction before calls, with tests on realistic data.
- Day 4Guardrails, data boundaries in prompts, strict output validation, key hygiene.
- Day 5A small evaluation set in CI, the report and the handover.
Evidence from the demo
The evidence below comes from my own demo project with synthetic data. It shows exactly what you receive; it is not a story about a client.
- IncidentScribe removes 13 kinds of personal data and secrets before calling Amazon Bedrock; in its 30-case evaluation, no case let raw personal data into a prompt.
- A hard cap ($0.15 a day, $4.50 a month) is checked before every call; when it's reached, users still get a rule-based summary.
- The free Bedrock Bill-Shock Calculator shows a typical support chatbot costing 8.7 times its pricing-page estimate, and which changes cut it most.
Demo project: IncidentScribe AI
What you keep
- Cost per feature, per day
- Caps with a fallback
- Redaction with tests
- Guardrail and output checks
- An evaluation set in CI
Not included
- Building new AI features
- Legal advice on privacy law
Questions
Do you need access to my production systems?
No passwords, ever. Audits use a read-only role you create and can delete at any time; changes arrive as a pull request you review and merge yourself.
How long does a project take?
2 to 10 working days depending on the package. You get the timeline and the fixed price before we start.
What if you find nothing worth fixing?
Then you hear that early, with the evidence. For the AWS Bill Doctor on accounts above $1,000 a month: if I don't find at least my fee in yearly savings, you don't pay.
What happens to my data?
Reports stay between us. When the project ends I delete your data and confirm it in writing.
How do I pay?
Upwork, international bank transfer, or UPI / bank transfer in India. Direct clients pay 50% upfront and 50% on delivery.