Ready to build AI into your business?
For companies building AI into their product or processes: ten questions on data, trust and production before the first model is chosen.
Our method
Trusted systems,
not AI demos.
It is how we build RAG systems, AI agents, data integrations and revenue diagnostics, and every project goes through the same five stages: whether it is a website that stopped earning, an AI product for a sales team or an assistant for lawyers and clinicians. The model comes fourth. Most of the risk, and most of our work, comes before it.
- Ingestion
- Normalization
- Trust
- Value
- Production
The five stages
From raw data
to a system you can act on.
Ingestion Start from the real data.
We bring in the data from where the business actually runs: CRMs, call transcripts, documents, spreadsheets, server logs, analytics, practice systems. Not a sample, not a demo dataset.
In practice: REST and SOAP/XML APIsOAuth2WebhooksCRM integrations (HubSpot, Salesforce)Call transcriptsDocuments and PDFsSpreadsheetsServer logsGA4 and Search ConsoleETL pipelinesRead-only access
Seen in: Affiliate Site Turnaround in 10 Weeks · AI Sales-Coaching Product, Working in 8 Weeks
Read the post →Normalization Make it mean the same thing.
Every source has its own idea of a customer, a note or a date. We turn them into one consistent, validated shape, because sophisticated AI on unreliable context still gives unreliable answers.
In practice: Data modelingTyped models (Pydantic)Schema validationEntity resolutionDe-duplicationID mappingTime zones and timestampsChunking and metadata for retrievalPostgreSQL
Seen in: The Data Layer Behind a Practitioner AI Platform · AI Sales-Coaching Product, Working in 8 Weeks
Read the post →Trust Make it safe to act on.
Structured outputs, validation, rules alongside the model, fallbacks, provenance, tests and human review where a decision is made. A successful request is not a correct answer.
In practice: Structured outputs (JSON Schema)Tool callingLLM evaluation (evals)Regression testsDeterministic fallbacksGuardrailsHuman-in-the-loop reviewProvenance and prompt versioningAudit trailsPII handlingHIPAA, SOC 2, GDPR
Seen in: AI Sales-Coaching Product, Working in 8 Weeks · A Live Affiliate Site in Three Weeks
Read the post →Value Solve a problem worth money.
Only now do we decide what the AI does: score, recommend, write, diagnose, automate. The goal is never AI adoption; it is a measurable outcome.
In practice: RAG (retrieval-augmented generation)EmbeddingsVector databasesHybrid searchAI agentsWorkflow orchestrationMCPScoring and classificationContent generationModel selection by measured quality
Seen in: Affiliate Site Turnaround in 10 Weeks · SuperEscuela: A Free School Built and Written with AI
Read the post →Production Survive reality.
Deployment, monitoring, failure handling, security, tests and feedback from real usage, with one team responsible from the first question to the running system.
In practice: CI/CD (GitHub Actions)ObservabilityStructured logging and correlation IDsRetries with backoffCircuit breakersRate limitingCloud deployment (GCP Cloud Run, Vercel)Cost and latency controlSecurity and data separationFeedback loops from real usage
Seen in: The Data Layer Behind a Practitioner AI Platform · SuperEscuela: A Free School Built and Written with AI
Read the post →
People and AI
AI speeds the work.
People stay responsible.
We use AI throughout analysis and development, because it makes levels of investigation, comparison and iteration practical that used to be too expensive. But we don’t treat model output as truth. Models produce false positives, jump to conclusions and optimize the wrong thing. So the human role moves up, and stays there:
- Understand
- Analyze
- Architect
- Plan
- Direct
- Validate
- Correct
- Deliver
AI accelerates the work underneath that loop. It doesn’t replace accountability.
The stack behind it
Tools we have shipped with,
chosen per problem.
We pick tools for the problem, not the other way round. These are the ones we have used in production.
- AI and LLMs
- Claude (Anthropic)OpenAIPrompt engineeringStructured outputsTool callingAI agentsMCPLangChainLangGraphLLM evaluation
- Retrieval and knowledge
- RAGEmbeddingsVector databases (Pinecone)ElasticsearchHybrid searchKnowledge systems
- Data and integration
- PythonFastAPIPydanticRESTSOAP/XMLOAuth2WebhooksPostgreSQLSupabaseETL pipelines
- Product and production
- TypeScriptReactNext.jsAstroVercelGCP Cloud RunGitHub ActionsTemporalObservability
- Marketing and revenue data
- GA4Google Search ConsoleGoogle Tag ManagerAhrefsTechnical SEOCore Web VitalsHubSpotSalesforceMarketo
- Platforms
- Drupal 7–11WordPressPHPSymfonySolr
What we won’t build
Some things look like progress
and aren’t.
- A chatbot because AI is fashionable, with no problem behind it.
- An AI that controls everything, with no rules, checks or accountable person around it.
- A report that diagnoses the problem and leaves the fixing to you.
- A prototype that only works on the demo data.
- A system nobody on your side can understand, run or maintain.
The method in depth
A short series on the blog takes each stage and the project that taught us the most about it.
Start the series →Have a problem
worth solving properly?
Tell us where your data lives and what you need it to do. We will tell you which stage is holding you back, and what the next practical step could be.
Prefer to write directly? enable JavaScript to see the address