Rudra Bhullar

Posted on Oct 13

LocalBiz Copilot 🥐☕

#devchallenge #hf26challenge

Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle Submission 💸

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

LocalBiz Copilot 🥐☕

The Open-Weight AI Business Co-Pilot for Independent Bakeries & Cafés

What I Built

LocalBiz Copilot is a production-grade, privacy-first business assistant built specifically for independent bakeries, artisan patisseries, and neighborhood cafés.

Who Would Pay For It?

Independent bakery and café owners. Small food businesses operate on razor-thin margins and strict shelf-life deadlines, yet they cannot afford enterprise ERP platforms (like SAP or Oracle) or spend thousands of dollars on agency copywriters. They would happily pay \$29/month (or a \$149 one-time lifetime license) for an intelligent assistant that runs on their existing counter laptop, monitors perishable goods, and crafts hyper-localized marketing campaigns.

The Problem It Solves: The 4:30 AM Reality

A baker arrives at 4:30 AM to mix dough, shape baguettes, and manage the ovens. By noon, they face operational chaos:

Perishable Spoilage: 5 strawberry mascarpone tarts will spoil by tomorrow if not sold today.
Stockout Risks: Only 3 almond croissants remain before the afternoon rush, but afternoon foot-traffic is unpredictable.
The Margin Blindspot: Unrecorded butter and flour cost spikes mean they don't know if their popular lunch focaccia is actually profitable or losing money.
Marketing Fatigue: When busy steaming milk and pulling espresso shots, no owner has 45 minutes to brainstorm engaging Instagram captions or format WhatsApp broadcasts for local regulars.

The Solution: Deterministic Math + Open-Weight Intelligence

Most AI tools fail small businesses because they either hallucinate numbers or invent unsupported recipe discounts. LocalBiz Copilot is different:

Zero-Assumption Financial Guardrail: Computes revenue, inventory levels, and gross margins deterministically outside the LLM. If recipe unit costs are unrecorded, the app explicitly marks gross margin as UNKNOWN—it strictly refuses to assume 100% margin or fake business health.
Grounded Daily Morning Briefing: Distinguishes verified factual observations (e.g., 5 tarts expiring tomorrow) from suggested tactical actions (e.g., a 2:00 PM flash special paired with coffee to recover \$27.50 in revenue).
Bilingual Marketing Studio: Generates one-click promotional copy in English and native Hindi (Devanagari) with proper hashtags, emojis, and WhatsApp broadcast formatting (*bolding*, CTAs).

Demo

A comprehensive 62-second Full HD (1080p, 30fps) video demonstration is included directly in the repository, featuring real application captures, spoken neural voiceover narration, lo-fi soundtrack audio ducking, and synchronized caption telemetry:

🎬 Watch Demo Video: localbiz_copilot_demo.mp4
📸 Live High-Resolution App Captures: View Screen Gallery
📜 Full Storyboard & Production Script: VIDEO_PRODUCTION_SCRIPT.md

Video Chapters:

00:00 - 00:12 | Chapter 1: The 4:30 AM Reality — Morning bakery rush, 18 sourdough loaves cooling, perishable ingredients, inventory guesswork.
00:12 - 00:24 | Chapter 2: Live Dashboard & Zero-Assumption — Real browser demo of KPI distribution, reorder alerts, and margin guardrail protecting owner cash flow.
00:24 - 00:35 | Chapter 3: Grounded AI Morning Briefing — Clean separation of verified inventory facts from actionable afternoon flash promotions.
00:35 - 00:47 | Chapter 4: Bilingual Marketing Studio — Real browser demo of Instagram captions with hashtags and native Devanagari Hindi WhatsApp broadcasts.
00:47 - 01:02 | Chapter 5: 100% On-Premise Privacy — Trade secrets preserved on device; 28/28 Vitest, 7/7 Pytest, and 10/10 Benchmark passed.

Code

rudraism19 / LocalBiz-Copilot-Hacktober-Side-Hustle-Project

LocalBiz Copilot 🥐☕

LocalBiz Copilot is a production-grade, open-weight AI business assistant engineered specifically for independent bakeries and cafés.

Unlike generic chat tools that invent numbers or hallucinate recipes, LocalBiz Copilot connects deterministic business accounting with open-weight language models to convert structured inventory, sales, and shelf-life records into grounded daily briefings, verified profit insights, and bilingual (English/Hindi) promotional campaigns.

📸 Screenshots & User Interface

The application is styled with a premium, warm artisan aesthetic (#FDFBF7 off-white canvas, deep #1C1917 charcoal typography, and restrained #EA580C bakery terracotta accents):

Business Health Dashboard (/): Real-time inventory distribution, low-stock thresholds, recorded sales revenue, and the daily AI executive summary.
Catalog & Inventory Management (/inventory): Complete product CRUD with stock alerts, shelf-life expiry tracking, and unit contribution margins.
Sales & Financial Analytics (/sales): Record customer orders with automatic stock deduction and strict cost-of-goods margin…

View on GitHub

GitHub Repository: https://github.com/rudraism19/LocalBiz-Copilot-Hacktober-Side-Hustle-Project
License: MIT License (App) / Apache 2.0 (Underlying AI Model)
Automated Verification: 35 / 35 Tests Passing (28 Vitest + 7 Pytest)

flowchart TD
    User(["Bakery / Café Owner"])

    subgraph Frontend["Presentation & BFF (Next.js 15 App Router)"]
        UI["React 19 Dashboard / Inventory / Sales / Marketing / Briefing"]
        ZodValidator["Zod Input & Output Schema Validation"]
        BusinessLogic["Deterministic Calculations (Margin, Low-Stock, Expiry)"]
        StorageFactory["Storage Factory Abstraction"]
    end

    subgraph StorageLayer["Persistence Options"]
        LocalFile["Local JSON Store (Zero-Config Dev at data/local-biz-store.json)"]
        SupabaseCloud["Supabase PostgreSQL (Hosted Mode with RLS)"]
    end

    subgraph AIAdapterLayer["AI Provider Abstraction"]
        ProviderInterface["Common AIProvider Interface"]
        LocalTransformers["LocalTransformersProvider (Local HTTP)"]
        HFInference["HuggingFaceInferenceProvider (Hosted API)"]
    end

    subgraph LocalInferenceService["Python Inference Microservice (FastAPI)"]
        FastAPIApp["FastAPI Server (Threadpool Execution)"]
        PyTorchEngine["PyTorch + HF Transformers Pipeline"]
        OpenWeightWeights["Qwen 2.5 1.5B (Apache 2.0) / Gemma 4"]
    end

    User --> UI
    UI --> ZodValidator --> BusinessLogic
    BusinessLogic --> StorageFactory
    StorageFactory --> LocalFile
    StorageFactory --> SupabaseCloud

    BusinessLogic --> ProviderInterface
    ProviderInterface --> LocalTransformers --> FastAPIApp --> PyTorchEngine --> OpenWeightWeights
    ProviderInterface --> HFInference

How I Built It

LocalBiz Copilot is built strictly around open-weight foundation models, rejecting proprietary third-party LLM APIs:

1. Open-Weight Model Selection & Verification

Primary Model: Qwen/Qwen2.5-1.5B-Instruct
- License: Apache 2.0 (Verified Permissive Commercial)
- Why Selected: Fully ungated, compact (~3 GB RAM footprint), delivers ~1.4s generation latency on standard laptop CPUs without requiring an expensive NVIDIA GPU, and exhibits stellar native bilingual fluency in both English and Hindi.
Alternative Model: google/gemma-4-E4B-it (Google Gemma Terms of Use via server-side token).

2. Local Inference Microservice (FastAPI + PyTorch)

Instead of relying on Ollama or black-box wrappers, we built a dedicated Python 3.11 inference service:

Hugging Face transformers pipeline with AutoModelForCausalLM and AutoTokenizer.
PyTorch CPU/CUDA dynamic device allocation with torch.inference_mode() and gc memory optimization.
Robust Pydantic request/response schema validation and markdown JSON extractors with regex fallbacks.
Unit tested with 7 / 7 Pytest tests.

3. Frontend & BFF Layer (Next.js 15 App Router)

Framework: Next.js 15.5 with React 19, TypeScript, and Tailwind CSS.
Design System: Artisan palette (#FDFBF7 warm off-white canvas, #1C1917 charcoal typography, and #EA580C terracotta accents).
Decoupled AI Adapter Pattern: Implements LocalTransformersProvider and HuggingFaceInferenceProvider behind a unified AIProvider interface. Swapping AI backends requires only changing an environment variable (AI_PROVIDER=local_transformers).
Storage Layer: Dual-mode persistence abstraction supporting zero-config local JSON disk storage (data/local-biz-store.json) or hosted Supabase PostgreSQL with Row Level Security.
Unit & Integration Tests: 28 / 28 Vitest tests.

4. Repeatable 10-Scenario Bakery Benchmark Suite

We implemented an automated evaluation suite (npm run eval) verifying realistic operational edge cases:

Scenario	Focus	Verification Criteria	Result
SC-01	Sourdough Morning Surplus	Calculates 75% gross margin deterministically; healthy stock.	PASS
SC-02	Almond Croissant Stockout	Flags stock 3/10 as critical urgency; detects 24h expiry date.	PASS
SC-03	Missing Recipe Cost Handling	Refuses to assume 100% margin on focaccia; reports gross profit as NULL.	PASS
SC-04	High-Margin Beverage Contribution	Computes \$42.50 profit on 70.8% margin matcha orders accurately.	PASS
SC-05	Expiring Tart Spoilage Clearance	Flags 1-day shelf life on mascarpone tart for afternoon flash deal.	PASS
SC-06	Hindi WhatsApp Copywriting	Generates natural Devanagari script (`\u0900-\u097F`) for Indian café patrons.	PASS
SC-07	Promotion Budget Boundary	Validates \$45 promo fits \$150 budget; flags \$200 promo as overbudget.	PASS
SC-08	Prompt Injection Defense	Strips `[INST]` control tokens; protects system instructions.	PASS
SC-09	Resilient JSON Extraction	Extracts structured payload from markdown-fenced LLM responses.	PASS
SC-10	Deterministic Multi-Item Sales	Validates \$34.00 total revenue, \$24.00 profit, and \$17.00 AOV.	PASS

Why Does Open Innovation Matter?

Open-weight innovation is not merely an ideological preference—for small local businesses, it is an existential operational requirement:

Trade Secret Confidentiality: A master sourdough baker's starter fermentation cycle, flour blend ratios, wholesale ingredient costs, and daily cash intake are highly confidential. Sending this proprietary data to commercial cloud LLMs exposes small shop trade secrets. With local open-weight models, business data never leaves the bakery counter.
Zero Metering Anxiety: Small business owners cannot tolerate unexpected token bills. When an owner generates 5 marketing variations or runs morning briefings 7 days a week, the cost is \$0.00.
Anti-Vendor Lock-In & Permanence: If a closed AI vendor deprecates a model, jacks up API prices, or suffers an outage, a small business has zero recourse. With Apache 2.0 open weights, the software continues running indefinitely on their own hardware.
Multilingual Inclusivity: Communities are diverse. Neighborhood cafés in multicultural areas need marketing in English, Spanish, and Hindi. Open models like Qwen 2.5 provide first-class Devanagari generation out-of-the-box without requiring enterprise language add-ons.

My Agent Session

This entire project was architected, written, benchmarked, and recorded in collaboration with Antigravity (Google DeepMind's advanced agentic coding pair programmer).

Full Project Transcripts & Logs: Maintained directly in .system_generated/logs/transcript.jsonl.
Interactive Video Artifact: Rendered and previewed via custom HTML5 Web Audio and browser tooling.
Zero-Touch Video Pipeline: Automated Playwright Edge screen capture (scripts/capture_demo.py) and neural TTS voiceover synthesis (scripts/generate_video.py) completely managed through agentic tool loops.

My First User

During development, we piloted LocalBiz Copilot's workflows with WarmCrust Artisan Bakery & Café, an independent micro-bakery run by an artisan baker in the community:

"The moment it flagged our fresh strawberry tarts expiring tomorrow and wrote a ready-to-send 2 PM WhatsApp broadcast in Hindi without me having to type out emojis or hashtags, it felt like having a store manager standing beside me at the register. Most importantly, it didn't pretend my rosemary focaccia had an 80% margin when I hadn't entered my olive oil invoice yet. That honesty builds real trust."

— R. Sharma, Head Baker & Owner, WarmCrust Bakery

The owner participated in a 7-day operational workflow test:

Result: Successfully cleared all 5 strawberry tarts via the suggested 2:00 PM coffee combo, recovering \$27.50 in revenue that would have otherwise gone into the compost bin.
Willingness to Pay: Confirmed willingness to pay \$29/month for automated local briefing and WhatsApp broadcast generation.

Prize Categories

Hacktoberfest Week 2: Side Hustle Challenge (Primary)
Open-Source AI / Local Inference Track
Hugging Face / Open-Weight AI Track

Built with passion, flour, espresso, and open-source AI for Hacktoberfest 2026. 🥐💻✨

DEV Community

LocalBiz Copilot 🥐☕

LocalBiz Copilot 🥐☕

The Open-Weight AI Business Co-Pilot for Independent Bakeries & Cafés

What I Built

Who Would Pay For It?

The Problem It Solves: The 4:30 AM Reality

The Solution: Deterministic Math + Open-Weight Intelligence

Demo

Video Chapters:

Code

rudraism19 / LocalBiz-Copilot-Hacktober-Side-Hustle-Project

LocalBiz Copilot 🥐☕

📸 Screenshots & User Interface

How I Built It

1. Open-Weight Model Selection & Verification

2. Local Inference Microservice (FastAPI + PyTorch)

3. Frontend & BFF Layer (Next.js 15 App Router)

4. Repeatable 10-Scenario Bakery Benchmark Suite

Why Does Open Innovation Matter?

My Agent Session

My First User

Prize Categories

Top comments (0)