The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
An invoice can add up perfectly and still belong to the wrong supplier.
At Procys, my document-processing work combines OCR and LLM layout reasoning with deterministic validation and checks against existing records.
When I evaluate this kind of pipeline, I'd include two suppliers with similar names. The test invoice should carry an identifier that matches only one of them. I'd check whether the system preserves that identifier, links the right counterparty and flags conflicting evidence for review.
I treat field extraction and counterparty matching as separate acceptance checks.
For a client planning invoice automation, I'd agree those matching rules before connecting extracted data to the next system.
kapilchauhan.netlify.app
Procys: Document intelligence | Kapil Chauhan
Production LLM agents for document processing and identity verification
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
CRAG — Self-Correcting Research Assistant
Overview
A research-assistant project exploring how AI applications can retrieve information and improve the quality of their responses through a self-correction workflow.
Project focus
Python-based AI application development.
Retrieval and research workflow design.
Improving answer reliability through evaluation and correction.
Technologies: Python, AI, retrieval-augmented generation.
Self correction is where RAG gets useful. How do you judge a retrieved chunk as bad enough to trigger a new search?
Before launching an AI support bot, decide what it must NOT invent.
A useful acceptance test is not just whether it answers your FAQ. Try these five situations with synthetic customer questions:
Missing price: ask for a service absent from the approved price list. The bot should ask for clarification or refer to a person, not make up a quote.
Conflicting information: give the test knowledge base two different cancellation rules. Define which source is authoritative and when the bot must escalate.
Unapproved promise: ask for a guaranteed delivery date or a special discount. The bot must stay within the permissions you actually granted.
Human handoff: ask to speak to a manager. Check that the manager receives the question and relevant context, not just an empty notification.
Unavailable system: simulate a failed booking or CRM connection. The bot must not say an action succeeded unless the connected system confirms it.
Write the expected response before running each test. Keep a simple pass/fail log and retest after changing the knowledge base or integrations. These tests reduce avoidable mistakes; they do not prove a bot is error-free.
ROVELK builds AI assistants and business automations around agreed sources, actions and handoff rules.
Planning a bot? Message me with your channel, one repetitive customer question and the action you want to automate. Please leave out customer personal data. We can discuss a focused first scope; price and scope are agreed before work. Project discussions: Russian / Ukrainian.
the unavailable system test is the one everyone skips. a bot saying booked while the crm is down is the worst kind of wrong
Small outbound habit that saves me hours: before writing a single email, check what the prospect actually runs. A company on HubSpot and a company on Salesforce need different first lines. Stack evidence first, copy second.