AI forecasting intelligence layer
An LLM layer on top of existing Snowflake forecasts that explains, in plain language, how external events affect confidence in fuel-consumption forecasts.
Outcome Adopted to help stakeholders interpret forecast confidence and external drivers.
- AI/LLM
- Snowflake Cortex
- Forecasting
- Solutions Engineering
- Snowflake Cortex
- LLM prompt orchestration
- Python
- SQL
- Teams and email integration
Context
A national logistics enterprise relied on fuel-consumption forecasts for operational planning, but stakeholders had no fast way to understand why a forecast might be less reliable in a given period. Fuel supply volatility and broader macroeconomic uncertainty could move the numbers, and leadership wanted that context surfaced alongside the forecast rather than discovered after the fact.
The ask was less “build a better forecast” and more “help people trust and interpret the forecast we already have.”
My role and ownership
I owned the proof of concept end to end: stakeholder discovery, feasibility, the build, and the demo, with support from colleagues and mentors. I worked directly with operations and data stakeholders to scope what “useful context” actually meant, then designed a solution that sat on top of the existing forecasting outputs rather than replacing them.
Constraints I worked under: the explanations had to be grounded in real data (no hallucinated drivers), they had to be concise enough for a non-technical audience, and they had to land in the channels people already used.
Technical approach
- Built an intelligence layer on Snowflake Cortex that read existing forecast outputs and relevant external-event signals.
- Controlled the LLM tightly with prompt templates and tool orchestration so the generated commentary stayed grounded in the underlying data and within a predictable format.
- Layered the data so the model explained external drivers (supply volatility, macroeconomic factors) against forecast confidence, rather than restating the forecast.
- Delivered the output into Microsoft Teams and email workflows so it reached stakeholders where they already worked.
- Iterated on prompt and context design through several rounds of stakeholder feedback.
Tradeoffs
The central tension was usefulness versus trust. A freer model produces richer narratives but risks inventing drivers; a tightly grounded model is safer but blander. I biased toward grounding and tight templates, accepting less narrative flair in exchange for explanations a logistics operator could rely on. I also kept the layer additive: it explained the existing forecast rather than introducing a second source of truth that stakeholders would have to reconcile.
Outcome
The layer gave executives and operations teams clearer visibility into forecast reliability and the external factors affecting it. The solution was adopted after the proof of concept and implementation discussions with Snowflake architects.
Reflection
The most valuable part was not the model, it was the framing: deciding that the problem was interpretation and trust, not prediction accuracy. That reframing is what made a lightweight, well-grounded layer the right answer instead of a heavier modelling effort. If I extended it, I would add a lightweight evaluation step so the quality of the generated explanations could be measured over time rather than judged by eye.