A use case to validate
An idea for an assistant, document search or automation exists, but its feasibility, quality and cost still have to be proven on your real data.
Expertise
AI and LLM integration,
working for
your product.
AI integration agency: we integrate language models (OpenAI, Anthropic, Mistral, self-hosted Llama) into your applications and processes, with RAG on your data, agents and multi-model routing, while keeping quality, cost and latency in check.
routeur.ts1// router.ts — illustration23export const routing = {4narrative: 'narrative model',5actions: 'instruction-tuned model',6search: 'RAG · vector database',7};89export const guardrails = [10'semantic cache', 'human oversight',11];
● QUALITY · COST · LATENCYOPENAI · ANTHROPIC · MISTRAL · LLAMA
Situations
What brings an AI project to us.
A prototype that does not hold up in production
The demo works, but answers vary, latency and the API bill keep growing, and nobody can measure quality.
Business data to make searchable
Documents, catalogues or internal databases must feed an assistant without leaving the company's trusted perimeter.
A hosting or confidentiality constraint
Data residency or control over hosting rules out some APIs and points towards self-hosted models.
Scope of work
What we integrate.
From feasibility study to operation: you can entrust us with the full integration or a specific component.
Discovery and feasibility
Use case, available data, quality criteria, cost per request and the role of human oversight, agreed before the first prompt is written.
RAG on your data
Retrieval-augmented generation over your documents and business data: chunking, vector indexing, source citations and access control.
Agents and tools
Conversational assistants and agents that call your APIs and business tools, with bounded and logged actions.
Multi-model routing
Model selection by intent, expected quality and cost, commercial models or self-hosted Llama, with a semantic cache for recurring requests.
Quality evaluation
Representative test sets, answer quality measurement and regression tracking whenever a model or prompt changes.
Integration and operation
Integration into the mobile or web application, latency and cost monitoring, logging and procedures for drift.
Method
From use case to AI in production.
- 01
Scope the use case
What the AI must produce, for whom, from which data and with what acceptable error rate. Cost per request and target latency are set at this stage.
- 02
Measure on your data
A prototype evaluated on a set of real cases: answer quality, cost and latency. Next steps are decided on these measurements, not on a demo.
- 03
Integrate in increments
Integration into the product, guardrails, human oversight where needed, tests and continuous integration.
- 04
Operate and adjust
Quality, cost and latency tracked in production; prompts, routing and models adjusted as usage evolves.
Responsibilities
Who decides what.
A typical division of responsibilities, agreed for each engagement.
| Topic | Your team | Black Tide |
|---|---|---|
| Data | You grant access to the relevant data and decide what may leave | We work within that perimeter and document the data flows |
| Quality criteria | You confirm what a good answer is | We build the test sets that measure it |
| Model choice | You weigh cost, quality and hosting | We compare the options on your real cases |
| Human oversight | You decide where human validation is required | We build it into the user journeys |
| Production release | You decide the release date | We prepare, deploy and monitor |
Selected work
AI delivered in production, not just prototyped.
Meduz
A cascading LLM pipeline combining a fine-tuned narrative model and an instruction-tuned model, an intent-based router and a semantic cache that cuts API calls by 40%.
Read the Meduz caseFantasy Alley
Around the reading app, a Discord community with a virtual economy and a conversational agent connected to a vector database.
Read the Fantasy Alley caseEach case describes the work delivered, its scope and the current status of the product.
Budget and next steps
How an AI project is priced.
What discovery establishes
Use case, usable data, quality criteria, candidate models and target architecture. This is the basis for a quote.
What makes up the budget
Design and engineering time, data preparation and the expected level of evaluation, plus model API or hosting costs, estimated during discovery and set out in the quote.
Running costs
Cost per request is measured from the prototype onwards. Model routing and caching keep it in check as usage grows.
Join at the stage you need
Feasibility study only, a measured prototype, or full integration into your product.
Useful questions
Before starting an AI project.
Does our data go to a model provider?
It depends on the chosen model and how it is hosted. When data residency requires it, we recommend self-hosted Llama models. Data flows are documented during discovery.
Which model should we choose?
There is no single answer. We compare OpenAI, Anthropic, Mistral or a self-hosted model on your real cases for quality, cost and latency, and can combine several.
Do we need to train a model?
Rarely as a first step. RAG and good routing cover most needs. Fine-tuning is worthwhile for a very specific style or task, like Meduz's narrative model.
Can you add AI to an existing application?
Yes. We first examine the application, its data and its constraints, then integrate AI in increments without weakening what exists, as a project or embedded in your team.
Next step
Tell us about your use case.
What the AI should do, the data available and your hosting constraints: the starting point for useful discovery.