AI3 min read
LLM costs do not scale the way your intuition says
A feature costing cents in a demo can cost thousands a month in production, and rarely because of the price per token. Where the money actually goes.
Customer-facing AI · chatbots, assistants, LLM features
AI Your CustomersActually Talk To
A chatbot that answers from your own content, an assistant that qualifies and books, an LLM feature built into the product you already ship. Grounded in your data, with a human handover where it matters — not a demo that falls apart on the second question.
AI chatbot for your site and messengers
A chatbot that knows your product because it answers from your own content, not from whatever the model remembers. It says “I don’t know” instead of inventing, and hands the conversation to a person the moment the question outgrows it.
Get startedWhat’s included
- Answers grounded in your own content, with sources
- Handover to a human with the full conversation
- Site widget, Telegram, WhatsApp
- Tone of voice and guardrails on what it may say
AI sales assistant
The visitor who would have left without asking anything instead gets a conversation: what they need, which option fits, and the next step booked. Everything it learns lands in your CRM, not in a chat log nobody reads.
Get startedWhat’s included
- Qualification and needs discovery in dialogue
- Product, plan, and pricing recommendations
- Booking a call and handover to sales
- Write-back to CRM with the qualified context
ChatGPT and LLM integration
GPT or Claude inside the product you already have — generation, summarising, extraction, semantic search — shipped as a real feature with a loading state, a failure state, and a cost you can predict.
Get startedWhat’s included
- OpenAI / Anthropic API integration
- Prompt design, evaluation, and versioning
- Streaming UI and graceful degradation
- Token cost control, caching, and rate limits
AI agents and RAG on your data
When one answer is not enough: agents that take several steps, call your APIs, and work over your own documents, catalogue, or knowledge base — with traces you can inspect when the output is wrong.
Get startedWhat’s included
- RAG over your docs, catalogue, or knowledge base
- Tool use and multi-step task execution
- Vector search and retrieval-quality tuning
- Traces, evals, and regression checks
Stack
What we reach for when building AI into a product.
Claude · GPT
RAG · vector search
Agents · tool use
Next.js · Node · Python
Site · Telegram · WhatsApp
Evals · cost control
Common questions
The five that come up in almost every scoping call.
How is this different from business automation?
By who is on the other side. This page is AI your customers interact with — a chatbot on your site, an assistant in your funnel, an LLM feature in your product. Business automation is AI your team never shows anyone: the support inbox, document processing, lead routing, internal reports. Different buyer, different success metric, so they are deliberately two pages rather than one long one.
What stops it from making things up?
Three things, in order. It answers from your content rather than from model memory, so there is something to check the answer against. It is allowed to say it does not know and pass the question to a person — a refusal is a correct answer, not a failure. And the answers are tested against a set of real questions before launch and again after every prompt change, so a fix does not quietly break something that used to work.
Which model do you use?
Whichever fits the task — Claude and GPT are both strong, and they are strong at different things. The integration goes behind a thin layer of our own, so swapping the model later is a config change rather than a rewrite. Being locked to one vendor is a risk in a market where the best option changes every few months.
How long does it take?
A chatbot answering from your content is a 2-3 week build. An assistant wired into CRM and booking, or an LLM feature inside an existing product, depends on what it has to reach into — that is a scoping conversation, not a number I can put on a page honestly.
What does it cost to run?
Model usage is a real monthly line item, so it gets estimated before we build, not discovered after launch. Cheap turns go to a cheaper model, repeated questions get cached, and there is a hard ceiling so a traffic spike or an abusive user cannot produce a surprise invoice.