Daily recap for June 7, 2026

Today we shipped 5 new step-by-step tutorials for small-business AI. They cover HR policy compliance, insurance quote comparison, a GitHub MCP server for DevOps, a Plaid financial Q&A agent, and a secure Databricks sandbox. Each comes with a downloadable zip and a full walkthrough.

New tutorials

Cohere Document Pipeline for HR Policy Compliance

Small businesses waste hours cross-referencing PDF and Word policy documents to stay compliant with changing regulations. This tutorial builds a document processing pipeline that ingests HR documents, extracts key clauses with Cohere’s LLM, repairs malformed output into guaranteed-valid JSON, and feeds a searchable knowledge base with automated compliance gap reports.

Read the tutorial →   Download the code (zip)

Built with Cohere, Next.js, Express. 90 tests, 92.09% coverage.

Cohere Agent Mesh for Insurance Quote Comparison

Independent insurance agents spend hours manually entering client data into carrier portals and comparing coverages. This multi‑agent system uses specialized agents per carrier, orchestrated by @reaatech/agent-mesh, to query carriers in parallel, compare results, and recommend the best policy—all through a Cohere‑powered mesh with Langfuse tracing.

Read the tutorial →   Download the code (zip)

Built with Cohere, agent-mesh, Next.js, Express. 106 tests, 93.82% coverage.

AWS Bedrock MCP Server for GitHub Small Business DevOps

Small dev teams lose time navigating GitHub’s UI to answer questions across repos. This MCP server exposes safe GitHub tools (search repos, list issues, get pull requests) to AWS Bedrock agents. A firewall blocks destructive operations, and Langfuse traces every call—keeping your GitHub data read-only.

Read the tutorial →   Download the code (zip)

Built with Bedrock, Express, MCP. 103 tests, 96.17% coverage.

Cohere Knowledge Agent for Plaid SMB Financial Insights

SMB owners waste hours manually combing through statements for cash flow and vendor spend answers. This natural-language Q&A agent connects Plaid-linked bank accounts, indexes transactions into Qdrant, and uses a hybrid RAG pipeline with Cohere to answer questions like “What was my cash flow last month?”—with session continuity for follow-ups.

Read the tutorial →   Download the code (zip)

Built with Cohere, Plaid, Qdrant, hybrid-rag. 135 tests, 98.38% coverage.

Databricks Code Sandbox for Secure SMB Data Analysis

Non‑technical staff need ad‑hoc Databricks reports, but writing SQL manually risks errors and cost. This sandbox translates plain English into safe SQL and Python, dry-runs code in an E2B sandbox, enforces security policies with a firewall, and tracks LLM and warehouse spend—all with session continuity for multi‑step analysis.

Read the tutorial →   Download the code (zip)

Built with Databricks, OpenAI, E2B sandbox. 103 tests, 96.49% coverage.

Building blocks shipped

agents-md-kit

Linting, validation, scaffolding, and observability for AGENTS.md and SKILL.md files. All 9 packages received a patch release that resolves a CI security audit failure, restoring green builds. The CLI, linter, and validator are the most commonly used. Browse the agents-md-kit catalog →

hybrid-rag-qdrant

MCP server and observability packages for the hybrid RAG pipeline. @reaatech/hybrid-rag-mcp-server migrated to zod 4.4.3 to fix tool input schema signatures; @reaatech/hybrid-rag-observability upgraded OpenTelemetry SDKs to 2.7.1. Browse the hybrid-rag-qdrant catalog →

idempotency-middleware

The Firestore adapter bumped @google-cloud/firestore to v8.5.0, a major version jump that may affect consumers pinning v7. Browse the idempotency-middleware catalog →

llm-cost-telemetry

Core types and CLI for tracking LLM spend. @reaatech/llm-cost-telemetry upgraded zod to 4.4.3; @reaatech/llm-cost-telemetry-cli bumped commander to 14.0.3. Both are minor releases. Browse the llm-cost-telemetry catalog →

llm-judge-toolkit

Evaluation, calibration, and bias detection for LLM judgments. All 10 packages patched a security audit (PR #38) and updated tooling. The engine, calibration runner, and CLI are the workhorses. Browse the llm-judge-toolkit catalog →

Browse the full catalog at reaatech.com/products.