AgenticMesh2026!
I run an agency because people want to work with me on gigs and I find great ones!
Core trick
Agents reconcile subcontractor timesheets inside Snowpark Container Services, so Cortex models run on data that never leaves Snowflake.
Description
I run an agency because people want to work with me on gigs and I find great ones! So this will help me reduce the 4-6 hours of manually reconciliation of my subcontractors to get my invoices out to my prime... With AI!
README excerpt
# Timesheet Reconciliation Agent > Built with Claude Code. Powered by Cortex Code. Deployed as a Docker container inside Snowpark Container Services — runs entirely within Snowflake. Two layers of AI — one to build it, one to run it. Built for the [Llama Lounge Agentic Hackathon](https://cerebralvalley.ai/e/llama-lounge-agentic-hackathon). --- ## AI in this project This project is a dogfood story: **Anthropic's AI built a product powered by Anthropic's AI, deployed entirely inside Snowflake.** ### Layer 1 — Claude Code built the app The entire Next.js frontend, Snowflake stored procedures, SQL schema, and API routes were written by **Claude Code (claude-sonnet-4-6)** via an iterative pair-programming session. Zero boilerplate written by hand. The finished app is packaged as a Docker image and runs as a **React/Node.js container inside Snowpark Container Services**, with Snowflake's Secure Data Access keeping all queries within the platform perimeter. ### Layer 2 — Claude 3.5 Sonnet runs the extraction - Timesheet screenshots (JPG) are sent directly to `claude-3-5-sonnet` via `SNOWFLAKE.CORTEX.COMPLETE` with multimodal `TO_FILE()` — no OCR pre-processing - Invoice PDFs are parsed by `SNOWFLAKE.CORTEX.PARSE_DOCUMENT`, then structured by the same model - All AI inference happens inside Snowflake as a SQL stored procedure call — no external API, no Python runtime, no separate service --- ## The Problem Automating Business Processes with AI can be error prone without