This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle
As a guy who has worked in sales and generated revenue of 1.4 MILLION Rs IN MY FIRST MONTH of business, i can tell you that this product is going to change how traditional Businesses work! Because i know where the Gap is
What I Built
ReplyDesk is a local, privacy-first AI customer-reply assistant built for solo boutique merchants, home bakers, and thrift apparel sellers running their side hustles on Instagram DMs and WhatsApp.
(Do check the docs https://github.com/Rajratna-D/Reply-Desk/tree/main/docs)
The Problem
Micro-merchants in emerging markets spend 3 to 5 hours every single day answering repetitive, multi-lingual buyer queries:
- "Kitne ka hai?" (Price inquiry)
- "Is size XL available in lavender?" (Inventory check)
- "Can you courier to Pune by Friday?" (Delivery timeline and fees)
- "What is your return window?" (Exchange policy)
Delayed responses kill conversions: buyers message 3 sellers simultaneously and buy from whoever replies first. However, solo creators cannot afford $30 to $50/month SaaS platforms or unpredictable cloud API tokens. Worse, generic cloud LLMs regularly hallucinate non-existent discounts, misquote inventory, and leak buyer contact information to third-party servers.
The Solution
ReplyDesk runs 100% on the merchant's local laptop. It ingests the shop's actual product catalogue, pricing rules, and delivery policies into an on-device SQLite database. When a customer inquiry arrives, ReplyDesk drafts an accurate, polite reply in the customer's language and dialect (English, Hindi, Marathi, or Hinglish) within 2 seconds.
Nothing is ever dispatched blindly: the merchant reviews drafts in a streamlined inbox triage queue, edits if needed, and approves with a single click.
Customer Message (WhatsApp / IG)
│
â–¼
[ PII Sanitization ] --> Masks phone numbers, preserves Order IDs
│
â–¼
[ SQLite FTS5 Hybrid Search ] --> Retrieves exact products, prices & policies
│
â–¼
[ Local Gemma 2B via Ollama ] --> Drafts natural reply in buyer's language
│
â–¼
[ Deterministic Guardrails ] --> Zero-hallucination verification (Price & Stock)
│
â–¼
[ Merchant Review Inbox ] --> One-click human approval & dispatch
Who Would Pay For This?
Independent boutique owners, artisanal makers, home bakers, and social commerce solopreneurs,anyone who wants to make their business grow and acquire new customers as well as retain them**. They would pay a one-time fee of $29 to $49 (or a $5/month local sync fee) because:
- It returns 3+ hours of their day back to making products and fulfilling orders.
- It completely eliminates cloud API bills (zero per-token charges).
- It guarantees that a chatbot will never underquote a custom cake or promise out-of-stock sizes. And i might even give it out for free. ---
Demo
- GitHub Repository: https://github.com/Rajratna-D/Reply-Desk
- Architecture Documentation: https://github.com/Rajratna-D/Reply-Desk/blob/main/docs/ARCHITECTURE.md
- Engineering Build Log: https://github.com/Rajratna-D/Reply-Desk/blob/main/docs/BUILD_LOG.md
- Do check the Future Upgrades : https://github.com/Rajratna-D/Reply-Desk/blob/main/docs/FUTURE_UPGRADES.md
Split-Screen Interactive Workflow
- Customer Simulator: Test inquiries in English, Hindi (Devanagari), Marathi, and Hinglish.
- Owner Inbox Triage: Review AI drafts side-by-side with matched catalogue context, confidence scores, and safety flags.
- Guardrail Telemetry: Real-time visual alerts if a price mismatch or inventory discrepancy is detected.
- Catalogue Live Search: SQLite FTS5 full-text search across product variants and seller FAQs.
- Interactive Benchmarking: Built-in evaluation page to run reproducible test suites on-device against live models.
Code
Rajratna-D
/
Reply-Desk
Local-first, review-only AI reply assistant for Indian boutique sellers. Powered by Gemma via Ollama across English, Hindi, Marathi, and Hinglish. 100% offline.
ReplyDesk: Local AI Customer-Reply Assistant for Small Sellers
DEV / MLH Hacktoberfest 2026 Open-Source AI Challenge (Week 2: "Side Hustle")
Built with open-weight Gemma 4 Effective 2B and Gemma 2 2B, running locally on-device with zero cloud APIs, zero per-reply subscription bills, and complete customer privacy.
The Problem
Small independent merchants running boutique shops on Instagram and WhatsApp (apparel, home bakeries, handmade jewellery) spend 3 to 5 hours every single day answering repetitive buyer inquiries:
- "Kitne ka hai?" (Price check)
- "Is size XL available in blue?" (Variant and stock check)
- "Can you deliver to Pune in 2 days?" (Delivery policy)
- "What is your return policy?" (Return window)
Delayed replies directly cause abandoned carts and lost sales. Many buyers inquire in a mix of Hindi, Marathi, and Hinglish. Small sellers often reply in a rush with inaccurate numbers or inconsistent details. Generic cloud chatbots hallucinate non-existent discounts, invent uncatalogued inventory…
The repository is open-sourced under the Apache License 2.0 and features a modular Python architecture:
- engine/service.py: Core orchestration pipeline handling draft generation, fallback triggering, and telemetry logging.
- engine/guardrails.py: Deterministic safety and price-grounding guardrails (whitelisting customer-echoed figures and volume order multiples).
- engine/retrieval.py: SQLite FTS5 hybrid retrieval with multi-turn pronoun expansion (prior_queries) and persistent embedding cache.
- engine/importer.py: Robust CSV/JSON shop catalogue ingestion with schema validation and preservation of seller-saved FAQs.
- eval/run_eval.py: Automated 34-case offline benchmarking suite measuring price accuracy, script consistency, and safety adherence.
- ui/: Modern Streamlit dark-mode interface with split-screen customer and merchant workflows.
How I Built It
1. Open-Source AI Stack
- Foundation Models: Open-weight Google Gemma 2 2B (gemma2:2b) and Gemma 4 Effective 2B (gemma4:e2b).
- Local Inference: Ollama running locally on CPU or consumer GPU (4-bit quantization, consuming under 1.8 GB of RAM).
- Zero Cloud Footprint: Completely functional with Wi-Fi disabled. All prompts, context retrieval, and generation happen on localhost.
2. Architecture & Deterministic Guardrails
Large language models alone are prone to hallucinating numbers and policies. ReplyDesk sandwiches the open-weight model between deterministic software layers:
- Pre-Flight PII Redaction: Masks customer telephone numbers and email addresses before prompt assembly, while safely preserving alphanumeric order IDs and courier tracking codes.
- SQLite FTS5 Hybrid Retrieval: Fast full-text indexing queries product names, descriptions, tags, and shop policies. If pronouns are detected in conversational follow-ups (e.g., "Does it come in red?"), the engine contextually expands the search query using previous turns.
- Structured Instruction Prompting: Guides Gemma to output strict JSON (draft_reply, language, confidence, needs_check, reasoning).
- Post-Generation Numeric Verification: Extracts all numbers from the draft reply and verifies them against the ground-truth catalogue. Legitimate multiples (e.g., buying 2 shirts for 2x price) and customer-quoted figures are allowed; invented prices or unlisted discounts are instantly flagged with a red needs_check warning for merchant review.
Why Does Open Innovation Matter?
Open innovation is what made ReplyDesk possible. A closed-source API model (e.g., OpenAI or Anthropic) fails the side-hustle merchant on every dimension:
- Operating Cost: Closed APIs cost $0.20 to $0.40 per 1,000 inquiries ($50 to $100/mo). ReplyDesk costs $0.00 to run on the merchant's existing laptop using ~0.001 kWh of power.
- Customer Privacy: Closed APIs send customer addresses, phone numbers, and margins to third-party clouds. ReplyDesk is 100% on-device; private conversations never leave the shopkeeper's hard drive.
- Offline Reliability: Closed APIs break during internet outages or traveling. ReplyDesk is 100% offline with zero network roundtrips.
- Longevity and Control: Closed APIs are subject to forced deprecations, rate limits, and price hikes. Open-weight models can run locally for years without breaking.
Open weights turn AI from a luxury corporate subscription into an accessible local utility for independent shopkeepers.
My Agent Session
ReplyDesk was designed, tested, and polished using an agentic pairing workflow. The test harness was developed iteratively with automated regression checks:
- 45 automated unit and regression tests passing in 15 seconds (pytest).
- Live telemetry benchmarking recorded in eval/reports/latest_eval.md.
My First Users
To ground ReplyDesk in real hustle conditions, I tested the system against two actual small businesses:
They will pay me if made the Agent according to their needs and preferences.
-
Sandhya's Home Kitchen (An independent woman who has a talent of making food products):
- Catalogue: festive foods, modaks, tea cakes, eggless cupcakes, and multiple delicious dishes, and strict 24-hour advance notice policies.
- Result: Gemma 2B handled bilingual queries ("Do you have eggless chocolate cake and can you deliver to Aundh today?"), politely informing the customer of the 24-hour baking window while highlighting same-day pickup options. Saved ~2.5 hours of texting during busy baking mornings.
-
Honey's Sofa Shop (An experienced craftsmanship in making Really good Sofas):
- Catalogue: Custom built sofas, Pre-Built Sofas and 7-day exchange policies.
- Result: Resolved the user needs, and noting their requirements, return inquiries across Hinglish and Marathi without confusing inventory status or hallucinating free shipping below the policy threshold.
Their data is Private but i have some sample users in the repo feel free to test with that
Trust me guys this has a lot of potential, when i will deploy this for user with all the API keys for seller's business, it is going to play a really huge role in that business
I will be using the Partner technologies for these upgrades for sure.
Prize Categories
- Primary Challenge Track: Hacktoberfest Open-Source AI Challenge Week 2: "Side Hustle"
- Open-Weight / Gemma Track: Best use of open-weight Google Gemma models (Gemma 2 2B / Gemma 4 Effective 2B) running entirely on-device via local inference.
Top comments (0)