Mill: A Local-First AI Operations Assistant for Food Side-Hustlers

From Kitchen Chaos to AI Operations Hub

Selling prepared meals—fitness meals, weekly family bowls, keto packages—is one of the hottest side hustles right now. But behind every beautifully packaged meal box lies operational chaos. Developer @justcallmemarko's project Mill, submitted in the second week of the Hacktoberfest Open Source AI Challenge, aims to solve these pain points with a local-first AI agent.

He summarized three core problems facing food side-hustlers: raw ingredients silently eat 15–20% of net profit through purchasing, weighing, and cooking oil consumption; when last-minute orders come in (like suddenly needing 15 teriyaki chicken rice bowls on Friday), there's no quick way to determine if inventory is sufficient; and when four burners are going, the oven is preheating, and 50 meal boxes are waiting to be filled, there's simply no free hand to check spreadsheets or take notes. Mill's approach is to turn a laptop into an autonomous kitchen command center.

According to the GitHub repository and dev.to submission notes, Mill is a local-first, autonomous AI assistant for food side-hustlers, with core capabilities built around inventory auditing, order management, meal prep guidance, and AI conversation.

Core Features Breakdown

Mill's feature design is tightly coupled to the actual workflow of a food side hustle, and can be broken down into four main threads.

Real-time inventory and raw material audit ledger: Tracks bulk ingredients down to gram and milliliter precision, recording exact unit costs. Every manual adjustment, restock, or kitchen fulfillment deduction is written to an immutable audit trail.

Recipe testing lab and profit pricing: Calculates precise batch ingredient costs and suggests selling prices based on target profit margins (70%+).

Visual kanban fulfillment pipeline: Drag and drop orders between Upcoming, Preparing, and Completed columns. When an order card is dragged to the Completed column, the system automatically calculates the required recipe ingredients, deducts them from inventory, and writes to the audit ledger.

Proactive meal prep mode: Step-by-step checklists that automatically identify countdown timers (e.g., "simmer for 15 minutes"), with audio-visual alerts to prevent burning an entire batch during boxing.

Autonomous AI sous-chef powered by FastMCP: This is an AI assistant directly connected to the local SQLite database, backed by 16 typed FastMCP tools. You can tell Mill: "Check if I have enough inventory to make 8 honey soy chicken rice bowls," "Record a Saturday order for Sarah," "Reduce chicken breast inventory by 500 grams," and it will autonomously complete database transactions.

Technical Architecture and Implementation Highlights

Mill uses a locally deployed architecture with separated frontend and backend. The frontend uses React 19, TypeScript, Tailwind CSS, with @dnd-kit for accessible, physical-feeling kanban drag-and-drop, and React Query for data fetching, supporting light and dark themes. The backend is built on FastAPI and SQLAlchemy 2.0 operating on SQLite, with Server-Sent Events for real-time AI token streaming.

The tool protocol layer is FastMCP, exposing 16 typed tools mounted at http://127.0.0.1:8000/mcp. The AI agent engine is based on openai-agents, connecting to OpenAI-compatible endpoints. The key design: by pointing OPENAI_BASE_URL to http://localhost:11434/v1, you can run open-weight models via Ollama (such as gemma, llama3.2, qwen2.5), fully localized, with zero code changes.

There are three implementation details worth noting. First, the immutable audit ledger: When dragging an order to the completed column, Mill doesn't simply subtract a number from the database—it looks up each recipe's scaled ingredient lines, calculates deductions, and records an audit transaction (source code at backend/src/mill/services/orders.py). Second, FastMCP tool execution: The author cites community research on tool descriptions and accuracy, noting that explicit type hints combined with deterministic database operations prevent reasoning errors in small local models. Third, kitchen-oriented interaction design: Kitchen environments require large tap targets, high-contrast buttons, and zero clutter. Any meal prep step with a duration automatically parses the time and renders an interactive countdown.

FastMCP and the MCP Protocol: MCP (Model Context Protocol) is an open protocol proposed by Anthropic in late 2024 to standardize how large language models interact with external tools and data sources—think of it as a "USB-C port for the AI world." It defines how tools are declared, described, invoked, and how results are returned, enabling the same model to connect to different tool services seamlessly. FastMCP is a Python server-side framework based on this protocol, where developers simply declare functions with decorators and type hints to automatically generate tool descriptions conforming to the MCP specification. Mill leverages this framework to wrap 16 database operations into typed tools, allowing small local models to safely complete multi-step transactions like inventory adjustments and order creation without touching raw SQL, balancing flexibility with controllability.

Ollama and OpenAI-Compatible Endpoints: Ollama is an open-source model runner for local inference, supporting one-click pulling and running of open-weight models like Gemma, Llama 3.2, and Qwen 2.5 on a laptop, and by default exposes an HTTP interface on port 11434 that is highly compatible with the OpenAI Chat Completions API. This "interface compatibility" is key: any client code originally written for OpenAI can seamlessly switch to local models by changing OPENAI_BASE_URL to a local address. Mill's AI agent layer, built on the openai-agents framework, leverages exactly this feature to achieve "one-click cloud/local switching" while retaining streaming SSE output and other capabilities, letting users freely balance cost, privacy, and hardware performance without changing business code.

Why Local-First Matters for Food Side-Hustles

Regarding Mill's open-source and local-first positioning, the author gives four reasons oriented toward business sustainability.

Food side-hustle net profit margins are typically only 15–25%. Token-billed cloud APIs erode profits every time you ask about recipe pricing or check inventory, whereas running open-weight models locally through Ollama incurs zero ongoing API costs. Commercial kitchens, basements, and meal prep spaces often have terrible Wi-Fi signal, and cloud-latency-dependent assistants fail outright during peak prep hours—Mill runs 100% on the local machine. Side-hustlers collect customer delivery addresses, food allergy information, and health goals; storing these in local SQLite and processing them with local models ensures customer data never flows into third-party model training pipelines. Since Mill uses a standard tool-calling contract, sellers can freely choose Gemma, Llama 3.2, Qwen 2.5, or any open model suited to their laptop hardware—no vendor lock-in.

The project offers a two-step startup process: after cloning the repository, the backend seeds realistic inventory, recipes, and sample orders via uv run mill-seed, and starts the API service with uv run mill-api; the frontend uses bun install and bun dev to bring up the kanban. After that, you can drag orders at http://localhost:5173 to watch inventory auto-deduct, switch to the Preparation tab to start guided cooking timers and converse with a scope-limited AI sous-chef, or use the AI Assistant tab to have Mill execute multi-step database adjustments through FastMCP tool calls.

Token-Based Cost Model: Current mainstream cloud LLM APIs generally use "token-based billing," where both input and output text segments (roughly 0.75 English words or 1–2 Chinese characters per token) are billed separately, and the more complex the reasoning or the more conversation turns, the higher the cost. For operations assistants that frequently query inventory and repeatedly confirm recipe pricing, a single session often involves thousands of tokens, accumulating significant expenses over time. Local inference has near-zero marginal cost—only a fixed cost for the initial model weight download and electricity, with no further charges for subsequent calls. This is precisely the core economic motivation for food side-hustlers with only 15–25% profit margins to prefer running open-weight models locally.

A Noteworthy Side-Hustle Tool Paradigm

Mill's value goes beyond being just a food side-hustle app. It demonstrates a reusable product approach: embedding local-first AI agents into the actual operational workflows of vertical industries, using typed tools to constrain model behavior, immutable audit trails to ensure business credibility, and open-weight models to control cost and privacy. For side-hustle scenarios with thin margins, unstable network environments, and high data sensitivity, this combination fits real-world constraints better than general-purpose cloud assistants. The project is submitted to both the Hacktoberfest Week 2 main track and the "Best Use of Gemma" award.

Hacktoberfest and the Open Source AI Challenge Background: Hacktoberfest is a global open-source contribution event held every October by DigitalOcean since 2014, encouraging developers to submit Pull Requests to open-source repositories. In the past two years, with the rise of generative AI, a dedicated "Open Source AI Challenge" main track has emerged, requiring participating projects to actually use AI capabilities within open-source frameworks. In addition to the main track, there are themed awards—such as the "Best Use of Gemma" mentioned here, which encourages participants to build applications based on Google's open-source Gemma series of weight models. Mill's submission to both awards means it meets the evaluation requirements for AI application completeness while also demonstrating adoption of the open-weight ecosystem in its model selection, offering valuable reference for readers wanting to understand the evaluation criteria of such competitions.

One-line Summary

Open-source project Mill uses a local-first AI agent to help food side-hustlers manage inventory, orders, and meal prep.

Paragraph Summary

Mill is a project submitted by developer @justcallmemarko in the second week of the Hacktoberfest Open Source AI Challenge, targeting prepared meal side-hustlers. It addresses three major pain points—raw materials silently eating 15–20% of profits, difficulty judging inventory for last-minute orders, and being too busy in the kitchen to take notes—by providing a real-time inventory audit ledger, recipe profit pricing, kanban fulfillment pipeline, proactive meal prep mode with countdowns, and an AI sous-chef driven by 16 FastMCP tools directly connected to local SQLite. Technically, it uses React 19 + FastAPI + SQLite, with the AI agent based on openai-agents, capable of running Gemma, Llama 3.2, Qwen 2.5, and other open-weight models via Ollama. When orders are completed, recipe ingredients are automatically scaled and written to an immutable audit trail. The local-first positioning avoids three risks: token-billed cloud API costs, unstable kitchen networks, and customer health data privacy—with no vendor lock-in.