The Intelligent Production Assistant (IPA) is an AI-powered operational intelligence platform for upstream oil and gas, built entirely on Snowflake. Three autonomous detection agents — Sentinel, Guardian, and Fiscal — continuously monitor production physics, enforce HSE compliance, and flag cost variances. A Cortex Agent (IPA_AGENT) answers complex cross-domain questions in natural language by combining Cortex Analyst text-to-SQL with Cortex Search document retrieval — turning hours of diagnostic work into seconds.

In 2023, the upstream oil and gas industry lost an estimated $62 billion to unplanned downtime, with equipment failures accounting for 42% of lost production days.

In Q3 2024, a major Permian Basin operator experienced a cascading failure across 12 rod pump wells. The root cause — gas interference leading to fluid pound — was documented in a maintenance report filed 72 hours before the first failure. No one connected the dots.

Data exists. Expertise exists. But they live in silos — SCADA in one system, maintenance reports in another, financial actuals in a third. Engineers spend 60% of their time hunting for context instead of solving problems.

Production Engineer: Sensor data is disconnected from engineering knowledge. Diagnostic cycles take 4–6 hours for failures that should take 30 minutes — at $15,000/hour in deferred production per well.

HSE Manager: Safety permits don't talk to real-time hazard monitoring. A hot work permit issued for Zone B with H2S sensors spiking goes undetected until after the incident, at an average cost of $1.2M per OSHA recordable.

Asset Manager: LOE spikes lack operational context. Budget cycles become dominated by forensic accounting, with 3-week delays in variance explanations.

Operations Superintendent: Critical insights are buried in document noise. With 47 daily reports to read, 2.3 hours per day are spent reading reports instead of acting.

Each engineer sees a fragment. Root cause analysis takes days. Permits, sensors, financials, and maintenance logs exist in separate systems with no shared intelligence layer.

One question. Complete context. Seconds instead of days. SCADA time-series, unstructured documents, and financial actuals are unified in Snowflake, queried by autonomous agents operating in real time.

Well-RP-05 shows dynamometer load dropping to zero. The Sentinel agent detects the >10% deviation from baseline in TAG_HISTORY and generates an alert via AI_COMPLETE(). IPA_AGENT then surfaces both the live sensor reading and the failure analysis document in a single response: "What caused the rod pump failure on Well-RP-05?"

Zone-B H2S sensors spike to 15 ppm during an active hot work permit with a 10 ppm stop-work trigger. The Guardian agent detects the violation by querying live sensor readings against active work permits and escalates immediately — before an incident occurs.

Well-A10 shows a $50,000 chemical expense spike. The Fiscal agent detects the high-spend transaction in FINANCIAL_ACTUALS and raises an alert. IPA_AGENT then retrieves the Apex Chemicals invoice from the knowledge base and explains the variance in seconds: "What was the $50,000 charge on Well-A10?"

The IPA deploys as a five-page Streamlit application in Snowsight:

Ready to deploy autonomous operational intelligence for your oil and gas assets?

GitHub Repository →

The repository contains the complete deployment script, SQL setup files, Streamlit application, and sample data generating 187,000+ rows of synthetic time-series across wells, rigs, and sensors.