This Django app takes in information about daily foot traffic to shopping centers and uses LLMs to talk about trends and anomalies in the data in natural language.

The website offers several different LLMs as engines. Claude 3 Haiku, Anthropic's lightweight 2024 release, does surprisingly well for most inputs, so that is the default. The highest-quality output is subjective, but it seems that ChatGPT 4.5 Turbo is consistently a step above, but is quite a bit slower than Haiku.

This app does not put too much emphasis on economical token usage. The prompt engineering is a bit verbose, which comes at a cost of extra API usage and time, but for this toy application, the differences are quite small. At scale, it would be important to consider the relative costs of concise prompting compared to output quality.

• Brad Friedman - [email protected]
• Project Link: https://github.com/bradfriedman/placer-exercise