Getting an LLM to extract invoice data is the easy part. What happens when models disagree, totals do not reconcile, or confidence is too low? That is the challenge I explored while building an AI-assisted invoice and receipt review app. Here is how I built the pipeline: 1. Document classification Classifies uploads as invoices, receipts, or unsupported documents and routes them to the appropriate extraction path. 2. Dual-source extraction and reconciliation I integrated Azure AI Document Intelligence using its prebuilt-invoice and prebuilt-receipt models, alongside Azure OpenAI, to independently analyze each document. Document Intelligence remains the primary source. The merge logic preserves its values, uses LLM output to fill in missing fields, and flags conflicts and their sources for human review. 3. Deterministic validation I implemented rule-based checks for required fields, amount reconciliation, duplicate fingerprints, and EU VAT structure and checksum validation using python-stdnum. Fields with extraction confidence below 0.80 are flagged for review. 4. Constrained GL account suggestions Azure OpenAI suggests a general ledger account using structured output. The app validates the suggestion against a predefined catalog, preventing the creation of invented accounts and keeping approval decisions in human hands. 5. Human review Reviewers can inspect extracted values, sources, conflicts, and validation findings, correct fields, select an account, and approve or reject documents. The app can also draft supplier correction emails for human review. It does not send them automatically. Tech stack: React, TypeScript, FastAPI, Pydantic, Azure AI Document Intelligence, Azure OpenAI, and Azure Container Apps. Review data is stored in SQLite on an Azure Files mount, with the app configured for a single replica. The core principle: AI for interpretation, deterministic code for validation, and humans for consequential decisions. I have attached a walkthrough showing the pipeline from upload to review. I would love to hear your thoughts.

I build AI systems and automation projects and am looking for someone to handle client communication and bring in business. I will handle the technical side, from auditing workflows and designing solutions to building and delivering them. You will focus on connecting with businesses, understanding their challenges, spotting opportunities for AI and automation, and turning those conversations into projects. If you enjoy talking to businesses, know how to find opportunities, or already have potential clients who could use these solutions, DM me. Let us see if we could work together!

I have been thinking a lot about what we are actually giving an AI agent when we connect it to our systems. It starts innocently enough. Give it a tool to search a database. Then an API. Then email. Then maybe MCP. Then a few more permissions because the workflow needs them. At some point, the agent is no longer just answering questions. It has an identity, credentials, access to data and the ability to take actions. That is where I think agent security gets really interesting. How are we handling: • agent identities and short-lived credentials • permissions outside the model • read vs write access • prompt injection through external data • tool-call monitoring • human approval for sensitive actions • and, importantly, what happens when we need to stop the agent immediately? I also think we need to stop treating the prompt as the security boundary. “Do not access this” in the system prompt means very little if the architecture still gives the agent access to it. What part of agent security are you paying the most attention to?

Someone said, “AI doesn’t like coming back empty-handed.” 😂 How true has this been in your experience? I would love to hear!

I realized my AI OS doesn’t need more agents—it needs better handoffs I’m building a personal and business operating system called the Shan Command Center. I initially thought the goal was to create as many AI agents as possible, but I’m realizing that agents without clear responsibilities can create more confusion—not less. Here’s the team I’m currently planning: 🤖Dwayne — Chief of Staff Reviews my businesses, appointments, school responsibilities, and priorities to create my daily briefing. 🎨 Ivy — Content Manager Turns my ideas into content plans, scripts, graphics, and posts. ⚙️ Nova — Automation Manager Handles CRM workflows, lead follow-up, trigger words, and system alerts. 📊 Shan Command Center — Source of Truth Brings everything together so I can see what needs my attention without opening five different platforms. My biggest lesson so far: every agent needs a clearly defined input, responsibility, output, and handoff destination. Otherwise, you’re just creating several smart assistants that don’t know how to work together. For those who have already built an AI OS: do you use one central orchestrator to manage your agents, or do your agents communicate directly with each other?