CrumbDesk
An experimental order management project for the custom baking industry. It uses a local Gemma model for natural language order parsing together with a deterministic rule engine for business decisions, emphasizing that AI only handles understanding while humans retain decision-making authority.
Timeline (last 90 days)
Developer Rishabh Raj published an experimental project called CrumbDesk on Dev.to, using local Gemma AI plus a deterministic rules engine to handle custom bakery order changes
CrumbDesk was submitted to week two of the Hacktoberfest open-source AI challenge, in the Side Hustle track
CrumbDesk adopts a dual-track mechanism: the local Gemma model is only responsible for parsing natural-language orders into structured fields (such as weight, flavor, pickup time, special requests), and does not directly modify order status or trigger production actions
CrumbDesk's deterministic rules engine comes preset with hard constraints such as modification windows, ingredient loss thresholds, and capacity limits, and AI parsing results must pass rule validation before proceeding to the next step
Raj deliberately limited AI permissions for three reasons: local Gemma ensures data never leaves the device, deterministic rules make decisions traceable and explainable, and separation of responsibilities means model upgrades do not affect the stability of business processes
CrumbDesk's architectural pattern can be transferred to scenarios such as florists changing delivery addresses, restaurants changing menu quantities, and handicraft makers changing custom sizes; it applies wherever understanding customer intent and deciding whether to execute can be separated
CrumbDesk's code has been open-sourced in the repository linked from the Dev.to article
The combination of a local small model plus a lightweight rules engine is sufficient to handle medium-complexity tasks such as order parsing, and the rules engine can cover business risk-control needs with an extremely low amount of code
All Facts (8)
The combination of a local small model plus a lightweight rules engine is sufficient to handle medium-complexity tasks such as order parsing, and the rules engine can cover business risk-control needs with an extremely low amount of code
50%UnverifiedCrumbDesk's deterministic rules engine comes preset with hard constraints such as modification windows, ingredient loss thresholds, and capacity limits, and AI parsing results must pass rule validation before proceeding to the next step
50%UnverifiedRaj deliberately limited AI permissions for three reasons: local Gemma ensures data never leaves the device, deterministic rules make decisions traceable and explainable, and separation of responsibilities means model upgrades do not affect the stability of business processes
50%UnverifiedCrumbDesk's architectural pattern can be transferred to scenarios such as florists changing delivery addresses, restaurants changing menu quantities, and handicraft makers changing custom sizes; it applies wherever understanding customer intent and deciding whether to execute can be separated
50%UnverifiedCrumbDesk adopts a dual-track mechanism: the local Gemma model is only responsible for parsing natural-language orders into structured fields (such as weight, flavor, pickup time, special requests), and does not directly modify order status or trigger production actions
50%UnverifiedCrumbDesk was submitted to week two of the Hacktoberfest open-source AI challenge, in the Side Hustle track
50%UnverifiedCrumbDesk's code has been open-sourced in the repository linked from the Dev.to article
50%UnverifiedDeveloper Rishabh Raj published an experimental project called CrumbDesk on Dev.to, using local Gemma AI plus a deterministic rules engine to handle custom bakery order changes
50%