Skip to content
reaatechREAATECH
All postsrecap

Daily recap for May 23, 2026

Five step-by-step AI tutorials for small business — from self-hosted financial analysis to Shopify voice agents — landed today, along with 22 new building-block packages for evals, prompt control, and secret rotation.

RecapBot3 min read

Today we shipped five new step-by-step tutorials that small business developers can download and run locally — no cloud lock-in, no hidden costs. Each comes with a downloadable zip, a full walkthrough, and test suites exceeding 95% coverage. On the package side, we published 22 new npm modules across three families, giving you Git-like prompt versioning, RAG evaluation tooling, and zero-downtime secret rotation.

New tutorials

Ollama Code Sandbox for SMB Financial Analysis

Run custom Python and SQL models on sensitive financial data without ever leaving your infrastructure. This self-hosted sandbox uses Ollama to generate code from natural-language requests, executes it inside an isolated Daytona environment, and enforces session-level spend limits with automatic circuit breaking. If your accounting team needs on-premises data analysis with governance, start here.

Read the tutorial → Download the code (zip)

Built with Ollama, Daytona, Express — 56 tests, 95.27% coverage.

OpenAI Invoice Extraction for Xero SMB Accounting

Stop retyping supplier invoices. This pipeline reads PDFs and images with OpenAI’s vision model, extracts structured data, and directly creates bills in Xero. High-confidence extractions post automatically; low-confidence ones route to a human review queue. Budget engines keep LLM spend predictable, and the whole thing runs on Next.js 16 with full observability via Langfuse.

Read the tutorial → Download the code (zip)

Built with OpenAI, Next.js, Express — 123 tests, 100% coverage.

Cohere Voice Agent for Shopify Order Status Checks

Let customers call a phone number and ask “where’s my order?” in plain English. This voice agent uses Cohere to understand intent, Twilio for telephony, and Shopify for real-time order lookups. Circuit breakers protect against Shopify outages, per-call budgets prevent surprise charges, and a handoff path to Zendesk handles complex inquiries.

Read the tutorial → Download the code (zip)

Built with Cohere, Express — 111 tests, 100% coverage.

OpenAI Multi-Agent Handoff for E-commerce Order Handling

A Mastra-orchestrated system of three specialist agents — order status, modifications, and cancellations — handles common support requests without human intervention. Idempotent refund logic prevents double processing, an LLM cache returns instant answers for repeated queries, and a per-conversation budget cap keeps costs under a dollar. Built on Next.js 16 with Stripe and Supabase.

Read the tutorial → Download the code (zip)

Built with OpenAI, Mastra, Next.js, Express — 109 tests, 97.26% coverage.

Anthropic Document Pipeline for SMB Lease Abstraction

Property managers and small legal teams can pull key dates, rent amounts, and clause language from lease PDFs automatically. The pipeline uses Claude for extraction, retrieves similar clauses from a ChromaDB vector store for context, and gates results by confidence — high-confidence outputs auto-populate an abstract, while low-confidence items get flagged for attorney review. One API key and you’re extracting.

Read the tutorial → Download the code (zip)

Built with Anthropic, VoyageAI, Next.js, Express — 146 tests, 99.70% coverage.

New packages on npm

Prompt Version Control

A family of packages that give you Git-like versioning, tagging, and deployment for LLM prompts. The @reaatech/prompt-version-control-server provides a Hono API with checksummed versions, A/B traffic splitting, and eval-gated promotion. A typed client and CLI ship alongside it, plus an MCP server that lets any editor fetch production prompts. The shared types package underpins them all with 25+ Zod schemas.

Browse the family: Prompt Version Control

RAG Eval Pack

Everything you need to measure and gate the quality of your retrieval-augmented generation. @reaatech/rag-eval-metrics calculates faithfulness, relevance, and precision without LLM calls; @reaatech/rag-eval-judge runs LLM-as-a-judge with multi-model consensus. The suite orchestrator ties it all together with cost tracking, quality gates, and a CLI. An MCP server exposes the tooling to any AI editor.

Browse the family: RAG Eval Pack

Secret Rotation Kit

Zero-downtime rotation of API keys, database credentials, and TLS certificates. The core @reaatech/secret-rotation-core manages lifecycles with rollback and propagation verification. Pluggable providers cover AWS Secrets Manager, GCP Secret Manager, and HashiCorp Vault. A sidecar process exposes a REST API and Prometheus metrics for integration into existing infrastructure.

Browse the family: Secret Rotation Kit

Browse the full catalog at reaatech.com/products.

More on this topic

Comments

Sign in with GitHub to comment and vote.

Loading comments…