Saved 200+ weekly hours at Cooke Chile

The Challenge

Cooke Chile runs a large salmon aquaculture operation across dozens of sites and zones. Their teams generate enormous volumes of operational data every week: production metrics, compliance reports, anomaly logs, biological indicators, performance dashboards. Most of it lives in Excel files spread across departments.

The data was there. Getting answers out of it was the problem. Every operational question required someone to open the right spreadsheet, cross-reference others, and translate the result into something a manager could act on. Production asked one set of questions. Compliance asked different ones. Finance asked different ones again. Each one waited on the same small group of technical people.

Decisions were slow. The data team was a bottleneck. And the information was siloed by whoever happened to own a given file.

The Solution

We built Cooke a private multi-department AI portal that turns their operational data into something every team can talk to in natural language. No SQL knowledge required. No spreadsheet hunting. No waiting on the data team for routine questions.

The portal is organized around specialized bots per department. Production has a bot trained on production data. Compliance has its own. Operations has its own. Each bot knows its data, its vocabulary, and its workflows. An orchestrator on top routes questions to the right bot, or coordinates between them when a question crosses departments.

Under the hood, the system runs three distinct execution paths depending on the question: safe SQL for structured tabular queries, a sandboxed Python environment for advanced analytics, and retrieval-augmented generation (RAG) over their internal document corpus for written knowledge.

How It Works

1. Ingestion pipelineExcel and CSV files are uploaded through a guided wizard. A FastAPI worker auto-detects headers, validates the schema, and ingests the data into a normalized tabular store. Files are versioned and per-tenant isolated through Postgres row-level security.
2. Multi-bot orchestratorWhen a user asks a question, an orchestrator routes it to the right department bot, or coordinates across multiple bots when the question spans more than one area. Each bot has its own system prompt, glossary, and access scope.
3. Three execution pathsDepending on the question, the bot picks the right tool:
   • Safe SQL for direct tabular queries over their data warehouse.
   • Python sandbox (E2B Code Interpreter) for advanced analytics, joins, and computations that SQL can't express cleanly.
   • Document corpus RAG for written reports, compliance docs, and internal procedures.
4. Structured, traceable answersEvery answer comes back with the data it pulled, the calculation it ran, and a confidence signal. The team can audit what the AI did and verify the result. This was non-negotiable for compliance.
5. Security and access controlEvery query is scoped to the tenant and the user. Department bots see only their data. Audit logs track who asked what. Rate limits prevent abuse. Built on Supabase with row-level security enforced at the database layer.

How Cooke's Teams Use It

Different departments ask the portal different things. A few real examples:

• "What production anomalies happened in Zone 10 last week, and how does that compare to the previous month?"
• "Summarize the compliance status across all sites for this quarter."
• "Which sites are trending below their performance targets and by how much?"
• "Pull the biological indicators for site X and tell me if anything looks unusual."

Questions that used to take a data analyst 30 minutes to an hour now get answered in seconds. And the analyst is freed up to work on the actual hard problems.

The Results

• 200+ hours saved per week across operations, compliance, and management teams.
• 3,000+ conversations handled by the portal since launch.
• 99.8% answer accuracy verified against ground-truth data.
• Data analyst bottleneck eliminated for routine operational queries.
• Faster, more confident decisions across the business.

Why It Matters

Most "AI for business" projects fail because they treat AI as a chatbot bolted onto existing systems. We built Cooke something different: an AI-native operational layer that understands their data structure, knows their departments, and integrates with how their teams actually work.

The technical depth matters because the business depth matters. Three execution paths instead of one. Department-aware orchestration instead of a single chatbot. Auditable, traceable answers instead of black-box responses. That's what turns AI from a demo into infrastructure.

Cooke didn't buy a chatbot. They bought a way for every department to have an analyst on demand.