Local Gemma app maps teams to agents

This is less a chatbot than a structured consulting workflow simulator, and that’s the right framing for it. The interesting bet is that local open models can produce useful organizational reasoning when you force them through tightly scoped JSON stages instead of freeform chat.

• –The team-analysis step is the strongest part: it grounds “can we replace this team?” in sector, headcount, client type, and stack instead of generic AI hype.
• –The JSON-forced retry loop is a pragmatic fix for small-model drift and makes the outputs usable downstream.
• –Local-only inference is a real differentiator for sensitive company data and makes model A/B tests cheap and repeatable.
• –Gemma 3 4B should be enough to demo the pipeline, but the role-design stage will probably separate toy outputs from genuinely useful ones.
• –The saved history files make it easier to diff model sizes and see whether stronger models actually improve decisions, not just prose.