Project: Arctic Borderlands Climate Impact Scorecard

Industry: Canadian Government Sponsored Climate Change Research

Client: Rick Pollack, Ph.D. (Advanced Analytic Solutions) and Donna Larsen (Data Path Systems)

Service: Strategic AI Consulting

Summary

Dan Christ turned a static climate research spreadsheet into a natural language query tool for Advanced Analytic Solutions and Data Path Systems. The Arctic Borderlands Climate Impact Scorecard tracks weather, caribou, fish, bird, and berry data collected from Indigenous communities in Canada’s Northwest Territories between 2009 and 2013. End users can now ask plain-English questions and receive answers grounded directly in the underlying dataset, with the entire system running on controlled infrastructure with no cloud AI APIs involved.

The Problem

The scorecard held five years of carefully collected research data, but it was accessible only to people who already knew how to navigate a spreadsheet. Community members, tribal council staff, and funders needed a way to ask direct questions — “What are the major findings for caribou?” or “Was 2012 an unusual year for weather?” — and get an accurate answer without a statistics background.

Why Local AI, Not a Cloud API

The client needed reporting stability and security. Cloud-hosted models change behavior between versions, can be deprecated without notice, and route data through third-party infrastructure. For a multi-year public research project, that instability was a disqualifying risk. Dan scoped and delivered a system where every inference runs on infrastructure he controls directly, with no data ever leaving a client-owned or consultant-managed server.

What Was Built

• Data pipeline: A Python process reads the source Excel workbook and pre-computes trend directions and interpretations before the AI ever sees a question rather than asking the model to analyze raw numbers at query time. This solved both context-size limits and accuracy problems in one step.
• Local inference: A locally hosted Llama 3.1 8B model, served through Ollama, generates answers using only the pre-processed data provided with each request.
• API layer: A FastAPI web service handles incoming questions and returns grounded responses.
• Guardrails: The system prompt constrains the model to the scope of the dataset, so it declines out-of-scope questions rather than guessing or referencing internal data structure.

Outcome

In a live client demonstration, the assistant:

• Correctly identified caribou health trends across the study period
• Correctly named 2012 as the most anomalous weather year in the dataset
• Identified the species most affected by climate variables
• Appropriately declined to answer questions outside the scorecard’s scope, rather than fabricating a response

The client’s response was immediate approval to move forward with the project and its inclusion into a formal proposal submission.

Engagement Status

The proof of concept and production deployment are complete. The system is available for a live demonstration upon request.

Want to see it run against your own data? Let’s talk.