Services · AI Integration

AI integration that earns its keep.

Here's what Claude actually does in the software we ship — not what AI will transform for your business.

How we integrate Claude

Four patterns, each in production or direct development.

Peter Kibet is Anthropic's Claude Code Ambassador for Kenya — we build with Claude every day, and we wire it into client software where it does something a human or a simpler rule cannot. These are the patterns we use.

Vision in the field

AI disease detection in MkulimaOS.

A coffee estate with 26,000+ plants cannot afford to miss leaf rust early. Field workers photograph plants from their phones during routine inspection rounds. The image goes to Claude Vision via the MkulimaOS mobile app; Claude returns a structured diagnosis: threat identified, confidence level, recommended action, and block location tied to the GPS coordinates from the same submission.

Threats it catches: coffee leaf rust, coffee berry borer, antestia bug damage, and general stress indicators. The result is logged against the block record and surfaced to the estate manager dashboard before end of day. No agronomist visit required for first-level triage.

Claude in the CQRS pipeline

AI-categorized farm finance ledger.

Farm accounting is messy. A single M-Pesa transaction might be casual labor wages, agro-input purchase, or transport — and the description is whatever the payer typed. Tax season on a large estate without categorized records is chaos.

In MkulimaOS, every finance entry passes through a Claude call in the write-side of the CQRS pipeline before it is committed. Claude receives the raw transaction description, amount, and context (block, activity type, supplier if known) and returns a structured category code, subcategory, and a short normalized description. The read-side sees clean, categorized data. Reports and KRA submissions come out without manual cleanup.

Structured-output schemas for ERP data

Turning unstructured data into typed records.

Client businesses often have years of data in Excel sheets, PDFs, and WhatsApp message exports. Migrating that into a structured system is where most software projects lose time to manual data entry.

We build ingestion pipelines that send batches of unstructured records to Claude with a strict JSON schema as the expected output — member records, inventory counts, supplier invoices, historical transaction logs. Claude returns typed, validated objects that map directly to the Prisma schema. The pipeline validates, deduplicates, and imports. Days of data entry become hours of review.

AI-summarized weekly operations reports

Ops reports that write themselves.

Estate managers and business owners need a weekly picture of what happened — labor hours, expenses incurred, tasks completed, anomalies flagged. Writing that report takes time; reading raw data takes more.

We build scheduled jobs that pull the week's structured data from the database, pass it to Claude with a report template, and receive a natural-language summary with key figures highlighted, anomalies called out, and recommended follow-up actions. The report is delivered by email or WhatsApp before the manager starts Monday. No one assembled it manually.

The approach

Where AI earns its place, and where it doesn't.

We don't add AI because it's expected. We add it where the alternative is a human doing something tedious, slow, or error-prone — and where a wrong answer has a defined cost we can measure.

Every AI integration we ship has a fallback. Disease detection flags for human review; it doesn't auto-spray. Finance categorization shows a confidence score; low-confidence entries queue for manual sign-off. The system is always the authority. Claude is the fast first pass.

If you have a process that's slow because it involves reading, classifying, summarizing, or extracting data from text or images — tell us. Most of those are solvable with a well-designed Claude integration and a sensible schema.

Start an AI engagement

Tell us what process you're automating.

We'll tell you whether Claude belongs in it — and if so, exactly where. Most AI scoping starts with a 1-week Sprint engagement (from $4,500).

Book a discovery call See all services