alchemistchaos/tasteloop-v2-self-improving-slides

Taste loop is an experiment in generating presentation slides that follow a real set of brand design rules, and automatically fixing the slides when they break those rules. It uses a tool called cognee, which stores information as a knowledge graph, to remember past mistakes and the fixes that worked for them. The process works in four steps. First, an AI agent builds a slide as a self contained HTML file. Second, a headless Chrome browser renders that slide into an image. Third, the image or the code is checked against a list of thirty brand rules pulled from real brand guidelines. Fourth, if something is wrong, the agent fixes it, using the rule's text and, when memory is turned on, notes about past mistakes and proven fixes for that exact rule. The slide is then checked again. The project's headline result compares the agent with this graph memory against the same agent without any memory. On eight new slides with several rule violations each, the version with memory reached full compliance on seven of the eight slides in about 1.4 repair rounds, roughly 45 percent less repair work. The version without memory did not fully fix any of the eight slides within the same budget of attempts. The repository includes the finished presentation as a PDF, an interactive HTML deck you can step through with arrow keys, a one page report of the results, and an interactive view of the knowledge graph itself. Running it requires a fairly involved setup: Node 20 or newer, Python 3.12, headless Chrome, a logged in Codex CLI account, and Ollama for cognee's local language model. An alternate backend called Baseten is supported but optional and not used by default.