Service 01 / AI Integration
Wire AI into the product,
not into the slide deck
You already have a site or an internal system and want AI in it,
but nobody on the team has shipped one before.
That's the layer I do: timeouts, hallucinations, cost overruns — the engineering.
Who this is for
- ·You have a site or internal tool and want an assistant or Q&A in it
- ·Inbound forms and emails are read by hand and you want them triaged
- ·You want AI in the product but aren't sure the API bill will stay sane
- ·Someone built you a demo — it worked in the demo and broke in production
Does this sound familiar?
「We want an AI assistant but don't know where to start」
「Customers send dozens of emails a day and we answer each one manually」
「It makes things up. Ask about something we don't cover and it just invents an answer」
「Our API bill went up fivefold last month and nobody knows which calls caused it」
What you get
What I don't do
Stating the limits saves both sides more time than listing capabilities.
No model training
No fine-tuning, no training runs. I integrate commercial model APIs. If your problem genuinely requires a trained model, I'll tell you that directly.
No 100% accuracy guarantee
Anyone promising that is lying. What's achievable: confidence signalling, cited sources, and an explicit "I don't know" outside scope.
Not a content mill
Integration only. Bulk marketing copy generation isn't what I do — I'll point you elsewhere.
No project without a concrete problem
"We should probably use AI somewhere" — I'll help you find the actual use case first. If there isn't one, I won't take the work.
Timeline and pricing
Fixed pricing, not hourly. Under $2,200: 50% upfront, 50% on acceptance. Above: three instalments. Third-party costs (server, API, domain) are not included.
Why take my word
Estimate the cost yourself first
The AI cost calculator is built on real model pricing. Four inputs and you know the monthly ballpark before committing to anything.
Fallbacks are not theoretical here
During this site's rebuild the API key actually went invalid. The page didn't break — it fell back to a template path. You never see that in a demo; it shows up in production.
The boundaries above are real refusals
Those four "don't do" items have been said to paying prospects. A refusal usually means direction mismatch, not price.
Common questions
Which model?
OpenAI-compatible by default — DeepSeek, GPT, Qwen and others all fit. I'll compare them on your budget and latency needs during scoping. Models can be swapped without rewriting the integration.
What will the API bill look like?
Depends on volume and context length. At 100 calls/day with 2,000-token contexts, most models land in the low hundreds of RMB per month. I'll give you a capped plan before you sign.
Is my data used for training?
Commercial APIs generally don't train on your data, though terms differ by provider — I'll confirm for whichever you pick. If the data is sensitive, local or private deployment is possible, at a different price point.
Can I edit it myself after handover?
Yes. Prompts are separate, commented files. Logs are viewable. The handover recording walks through it. Only logic changes need me.
What if it doesn't work well?
Acceptance criteria (accuracy, latency) are set during scoping. If they aren't met, it isn't accepted and the balance isn't paid. Writing the criteria early protects both sides.
Just describe the problem
No requirements document needed. Tell me who does this task now, how slow it is, and where it goes wrong. I'll say whether AI fits and roughly what it costs.
Describe the problemOther services
Public pricing, written boundaries, and an honest no when it's not a fit. That saves more time than any promise.