Developer Builds Self-Healing AI Middleware for Odoo ERP, Documents Real Failures

A developer built an AI-powered middleware layer for the Odoo ERP platform that dynamically introspects live database schemas at runtime, eliminating the need for hardcoded field assumptions that typically cause integration crashes. The system uses semantic vector search to map incoming data to correct fields and employs a LangGraph state machine to autonomously detect, diagnose, and retry failed writes using a local LLM. Running entirely on local models with no external API dependencies, the system achieved an 83.3% self-healing success rate across a small labeled evaluation set, but only a 33.3% first-attempt success rate. Key failures included the local LLM ignoring explicitly provided schema context in favor of outdated Odoo knowledge from pretraining, and a semantic matching ambiguity between similarly named fields that correctly triggered human review rather than an automated guess. The developer documented these shortcomings as the most instructive outcomes, highlighting the real trade-offs of using smaller local models over hosted API alternatives.