Daily recap for May 14, 2026

Today we shipped three step-by-step tutorials for small-business AI and published 31 new packages underneath them. Each tutorial is a working project you can download, run, and adapt — not a whitepaper.

New tutorials

Mistral AI invoice reconciliation for Stripe

This tutorial builds an Express API that accepts PDF invoices, extracts structured data with Mistral’s vision model, and reconciles the totals against your Stripe transactions. Discrepancies are flagged in a report that gets emailed to your finance team. The pipeline uses confidence-based document classification, structured output repair, and a budget cap per file — so costs stay predictable even with messy real-world invoices.

Read the tutorial → Download the code (zip)

Built with Mistral, @reaatech/confidence-router, @reaatech/structured-output-repair, and @reaatech/agent-budget-engine. 75 tests, 97.55% coverage.

AWS Bedrock multi-agent handoff for Zendesk ticket triage

This tutorial sets up a Next.js route that listens for new Zendesk tickets, classifies the intent, and hands the ticket off to a specialized AI agent — billing or tech support — running on AWS Bedrock. Context follows the ticket so the agent picks up right where the user left off. Per-ticket budget controls keep LLM costs in check, and a full escalation path flags tickets that need a human.

Read the tutorial → Download the code (zip)

Built with AWS Bedrock, @reaatech/confidence-router, @reaatech/agent-handoff, and @reaatech/agent-budget-engine. 114 tests, 97.62% coverage.

Google Gemini reliability suite for SMB AI operations

This tutorial wraps Gemini-powered agents in a production reliability layer. Circuit breakers isolate failing tools so one bad call doesn’t crash your whole workflow. Idempotency middleware prevents duplicate orders, and automated incident runbooks kick in when errors cross a threshold. A budget engine auto-downgrades to cheaper models if spend gets high. The whole thing runs inside a durable workflow system so you don’t lose state on restarts.

Read the tutorial → Download the code (zip)

Built with Google Gemini, @reaatech/circuit-breaker-agents, @reaatech/idempotency-middleware, @reaatech/agent-runbook, and @reaatech/agent-budget-engine. 78 tests, 100% coverage.

Building blocks shipped

Idempotency middleware

Prevents duplicate execution of side-effecting operations by caching responses and enforcing distributed locking. The core @reaatech/idempotency-middleware ships with pluggable adapters for DynamoDB, Firestore, and Redis, plus Express and Koa bindings. Pick the adapter that fits your stack.

Browse the family →

LLM cache

Speeds up LLM-powered features by caching responses with both exact-match lookups and semantic similarity search. Adapters cover DynamoDB, Qdrant, and Redis. A cost tracker calculates savings by model, and an HTTP server exposes cache management over REST. Observability is baked in with structured NDJSON logging and Prometheus metrics.

Browse the family →

LLM cost telemetry

Tracks token usage and cost across OpenAI, Anthropic, and Google models. Wraps the official SDKs so you don’t change your code. The aggregation layer groups spend by tenant or feature and enforces budget thresholds. Exporters push data to CloudWatch, Cloud Monitoring, or Loki, and there’s an MCP server for Claude Desktop or Cursor integration. A CLI lets you run reports from JSON logs.

Browse the family →

LLM judge toolkit

Evaluates LLM outputs with configurable templates for faithfulness, relevance, and safety. The engine handles retries, rate limiting, and caching. Consensus strategies combine scores from multiple models, and bias detectors flag position, length, or style preferences. Calibration tools measure against human labels with Cohen's kappa and confusion matrices. A CLI processes batch evaluations from JSONL.

Browse the family →

Browse the full catalog at reaatech.com/products.