DEV Community: Aditya

ShelfSense: Automated E-Commerce Stockout Arbitrage Engine with TabPFN, Gemma & ElevenLabs

Aditya — Sat, 17 Oct 2026 15:39:36 +0000

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

What I Built

Independent Shopify and Amazon marketplace merchants operate in a brutal, margin-compressed environment. When a high-volume competitor runs out of stock, customer search demand does not vanish—it instantly shifts to the next available merchant. Capturing this "stockout vacuum" window via dynamic price surges and targeted ad campaigns yields immediate, high-margin profit. However, monitoring competitors manually 24/7 and calculating inventory depletion curves from messy vendor CSVs is impossible for solo sellers.

ShelfSense is an autonomous inventory arbitrage engine that solves this end-to-end:

Forecasts Internal Stockout Runway: Uses TabPFN-3.5 to run zero-shot non-parametric regression directly on historical inventory velocity CSVs, determining exact remaining runway days without requiring custom dataset training or hyperparameter tuning.
Monitors Competitor Vacuums in Real-Time: Scrapes Google Shopping listings with SerpApi to detect competitor stockout flags and live pricing spreads.
Formulates Arbitrage Directives: Employs Gemma to reason across tabular runway metrics and live competitor shortages, generating surge pricing rules and 30-character search ad headlines.
Delivers Morning Spoken Briefs: Synthesizes a natural, 60-second executive audio briefing via ElevenLabs so merchants can listen to overnight market shifts during their morning routine.

Who would pay for this? High-volume Amazon FBA and Shopify D2C arbitrage sellers managing 20–500 SKUs. A \$49/month micro-SaaS tier is instantly justified if the engine captures just one competitor stockout window per month.

Demo

Live Application: https://self-sense.vercel.app
Live Backend API: https://selfsense.onrender.com
Swagger Documentation: https://selfsense.onrender.com/docs

The application features a Warm Industrial Gray brutalist design system (built with Next.js, Framer Motion, and Tailwind CSS) featuring a 12-column architectural wireframe grid, drag-and-drop CSV intake console, live competitor price matrices, and an integrated audio frequency wave player.

Code

The complete full-stack codebase—including the FastAPI backend service, TabPFN/SerpApi/Gemma/ElevenLabs orchestration pipelines, integration tests, and the Next.js frontend—is open source on GitHub:

AdityaNarayanPadhi / SelfSense

Autonomous D2C inventory arbitrage engine for Amazon & Shopify sellers. Forecasts stockouts with TabPFN-3.5, tracks competitor signals via SerpApi, reasons pricing rules with Gemma, and delivers daily morning audio briefs via ElevenLabs.

ShelfSense 🚀

E-Commerce Inventory Arbitrage & Stockout Detection Engine

ShelfSense is a high-throughput, production-ready backend service designed for high-volume Shopify and Amazon merchants. It predicts internal inventory depletion with state-of-the-art tabular foundation models, continuously crawls real-time competitor prices and stockouts, synthesizes dynamic pricing arbitrage playbooks with Google Gemma, and delivers a personalized 60-second morning audio brief synthesized by ElevenLabs.

🏛️ System Architecture

flowchart TD
    subgraph Client ["Client Layer"]
        FE["Next.js / Vercel Frontend"]
        Merchant["Merchant / Automated Job"]
    end
    subgraph ShelfSenseAPI ["ShelfSense Backend (FastAPI on Render)"]
        Router["/api/analyze Endpoint"]
        AudioRoute["/audio/{filename} Endpoint"]
        subgraph Engine1 ["1. Tabular Forecasting"]
            TabPFN["TabPFN-3.5 Regressor<br/>(Depletion Days & Velocity)"]
        end
        
        subgraph Engine2 ["2. Competitor Intelligence"]
            Serp["SerpApi Google Shopping<br/>(Live Prices, Stores & Stock Status)"]
        end

        subgraph Engine3 ["3. Reasoning & Arbitrage"]
            Gemma["Google Gemma LLM<br/>(Dynamic Pricing & Ad Copy)"]
        end

        subgraph Engine4 ["4. Audio Brief Synthesis"]
            Eleven["ElevenLabs API<br/>(eleven_multilingual_v2)"]
        end
    end

    Merchant -->|CSV + Product Query| FE
    FE -->|POST /api/analyze| Router
    Router -->|Async Gather|

…

View on GitHub

How I Built It

ShelfSense coordinates four specialized AI and cloud technologies into a non-blocking asynchronous pipeline:

Tabular Machine Learning (TabPFN-3.5): Small merchant inventory logs (typically 5 to 50 rows of historical sales velocity and price variance) are too small to train gradient-boosted trees or deep neural networks without extreme overfitting. We integrated Prior Labs' TabPFN-3.5 Regressor, passing the historical batch to predict exact days-to-stockout on the current inventory row.
Market Grounding (SerpApi): The backend asynchronously queries the Google Shopping engine via SerpApi, parsing merchant names (Amazon, Walmart, Best Buy), numerical price spreads, and out-of-stock badge snippets.
Cognitive Arbitrage Reasoning (Gemma): Using Google's open-weight Gemma, the engine receives the combined context: TabPFN's depletion runway and SerpApi's live stock status. If competitors are depleted, Gemma formulates a 10%–20% surge pricing recommendation and drafts high-conversion search ad copy.
Voice AI (ElevenLabs): The spoken segment from Gemma's directive is synthesized using ElevenLabs' eleven_multilingual_v2 model into a streamable MP3 brief.
Cloud Infrastructure:
- Render: Hosts the Python 3.11 FastAPI backend, handling async task execution, CSV parsing, and model coordination.
- Vercel: Hosts the Next.js App Router frontend with real-time audio playback and telemetry visualizations.

Why Does Open Innovation Matter?

Open innovation and open-weight models are what make ShelfSense commercially viable as a sustainable micro-SaaS:

Deterministic Arithmetic vs. Closed LLM Hallucination: Proprietary conversational LLMs notoriously struggle with tabular numerical calculations, often hallucinating stockout trends or failing at multi-variable regression. By utilizing TabPFN, we use a true tabular foundation model that performs Bayesian inference over tabular priors directly on CSV bytes with zero arithmetic drift.
Viable Unit Economics for High-Frequency Scraping: Running continuous competitor inventory audits across hundreds of SKUs with proprietary closed APIs leads to compounding per-token API bills that erode thin e-commerce margins. Open-weight models like Gemma allow continuous local or self-hosted batch inference on fixed-cost infrastructure (like Render) without unpredictable token markups.
Data Privacy for Sensitive Supplier Ledgers: Merchants are reluctant to upload proprietary supplier costs and sales velocity data to closed third-party model providers that may use commercial data for training. Open-weight pipelines allow running inference in controlled, private runtime environments.

My First User

Before finalizing the submission, I shared ShelfSense with a solo e-commerce seller managing home audio and electronics accessories across Amazon and Shopify.

The Test: They uploaded a sample 7-day velocity ledger for over-ear wireless headphones while a major category competitor experienced stockouts.
Their Feedback: "Normally, I only find out a competitor is out of stock days after the surge has already passed. Seeing the TabPFN runway countdown paired with the ready-to-run Google ad copy and price surge suggestion in an audio brief was surreal."
Willingness to Pay: They confirmed that an automated daily alert catching even two stockout arbitrage opportunities a month would justify a \$40–\$50/month subscription.

Prize Categories

Best Use of TabPFN (\$200) — Uses TabPFN-3.5 tabular foundation regression to forecast inventory depletion and categorize stockout urgency.
Best Use of Render (\$200) — Deployed the Python FastAPI backend service and AI runtime on Render using an automated build blueprint.
Best Use of Gemma (\$200) — Uses Google's open-weight Gemma model to analyze market pressure, formulate dynamic pricing rules, and generate search ad copy.
Best Use of SerpApi (\$100) — Uses SerpApi Google Shopping engine to crawl real-time competitor prices, store listings, and stock availability.
Best Use of ElevenLabs (\$100) — Uses ElevenLabs Voice AI to synthesize a daily 60-second morning audio brief for merchants.

FoliageRun — Autonomous Autumn Canopy Route Synthesizer for Bhubaneswar published: true

Aditya — Sun, 11 Oct 2026 18:25:14 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

FoliageRun is an autonomous running route synthesizer created to help developers and runners disconnect from screens and explore the outdoor canopy of Bhubaneswar, Odisha.

Urban runners frequently end up on high-traffic, polluted roads simply because standard GPS navigation prioritizes vehicular speed over greenery. FoliageRun solves this by orchestrating an open-source AI agent workflow that:

Ingests a runner's custom starting point, ending point, or distance target.
Maximizes tree-canopy coverage across major green corridors (such as Ekamra Kanan, Jayadev Vatika, and Chandaka trails).
Synthesizes realistic running loops and generates turnkey GPX 1.1 exports compatible with Garmin, Strava, and Apple Watch.
Provides a tranquil, distraction-free "Lock Screen" dispatch mode that encourages users to pocket their devices and touch grass.

Demo

Live Deployed App: https://foliagerun.vercel.app
GitHub Repository: AdityaNarayanPadhi/FoliageRun

Code

AdityaNarayanPadhi / FoliageRun

An open-source autumn running route generator for Bhubaneswar, Odisha. Built for the Hacktoberfest 2026 Touch Grass challenge.

🍂 FoliageRun Backend Service

Hacktoberfest 2026 "Touch Grass" Challenge Entry
An autonomous outdoor autumn running route generator powered by the Mastra framework, open-weight LLMs via Groq / Hugging Face Inference API, OpenStreetMap Overpass API, SerpApi, and telemetry tracked by Sentry AI Agent Monitoring.

🧭 Overview

FoliageRun encourages runners and outdoor enthusiasts to get outside and "touch grass" during peak autumn foliage. Given a starting point and desired distance (e.g., 5km, 10km), the backend orchestrates specialized agents and tools to synthesize a closed-loop running route that maximizes tree canopy coverage, routes through local parks, and includes live foliage condition reports and turnkey GPX export for Garmin, Strava, and Apple Watch.

🛠 Architectural Stack

Runtime & Language: Node.js (v20+ / v22+), TypeScript, Express
Agent Orchestration: Mastra Framework (@mastra/core)
Open-Weight LLMs: google/gemma-2-9b-it or meta-llama/llama-3.1-8b-instant via Groq API or Hugging Face Inference API using…

View on GitHub

The repository is fully open-source under the MIT License and includes:

Express + TypeScript backend orchestrator with Mastra Agent tooling.
Dynamic OSM Overpass query integration for parsing parks and green paths.
SerpApi tool integration for live foliage reports.
Sentry OpenTelemetry instrumentation for AI agent monitoring.

How I Built It

FoliageRun combines modern open-weight AI inference with real-time geospatial extraction:

Agent Orchestration (Mastra): Built with @mastra/core to coordinate external tools, handle context windows, and manage tool-calling lifecycles.
Open-Weight Inference (Gemma 2 9B): Powered by google/gemma-2-9b-it via Groq's high-speed endpoint. Gemma synthesizes optimal running waypoints, elevation estimates, and natural turn-by-turn briefings without proprietary closed-model lock-in.
Geospatial & Foliage Tools:
- OpenStreetMap Overpass API: Live queries identify paths, parks (leisure=park), and nature reserves (landuse=forest) around Bhubaneswar.
- SerpApi: Retrieves live seasonal canopy status, weather advisories, and park trail conditions.
Telemetry & Tracing (Sentry AI Monitoring):
- End-to-end tracing across ai.agent, ai.tool, and ai.chat_completions.create spans.
- Monitored latency, error rates, and prompt/completion token consumption directly within the Sentry dashboard.

Why Does Open Innovation Matter?

Outdoor well-being and navigation tools should not depend entirely on opaque, closed-source ecosystems.

Open innovation makes it possible to:

Pair community-maintained datasets like OpenStreetMap with open-weight models (gemma-2-9b-it), ensuring anyone can self-host and customize route synthesis for their own city without incurring per-query subscription fees.
Transparently inspect how route decisions are made, avoiding proprietary ad-driven route biases.
Give developers complete ownership over telemetry and data governance by pairing open-source toolkits with transparent observability platforms.

Prize Categories

Best Overall Touch Grass Project
Best Use of Mastra (Used @mastra/core for agent workflow execution and multi-tool orchestration)
Best Use of Sentry Agent Tracing (Integrated @sentry/node capturing automated spans for ai.agent, tool execution latency, and token consumption metrics)
Best Use of Gemma (Deployed gemma-2-9b-it for intelligent spatial waypoint reasoning and running briefing generation)