Storefront Ops Agent

Inspiration## Inspiration

Managing an online store's product catalogue is tedious, recurring manual work. Descriptions and tags get left blank, especially as a catalogue grows, and nobody checks consistently. I run Trendy Nest, a fashion brand, and this exact problem — incomplete listings quietly hurting sales and SEO — is something I've dealt with firsthand. I wanted an agent that doesn't just chat about the problem, but actually watches the catalogue and fixes what's safe to fix, without me having to remember to check.

What it does

Storefront Ops Agent audits every product in a Shopify store. For any listing with a missing description or missing tags, it uses Gemini to generate appropriate content and writes it back to Shopify — then re-reads the product to verify the write actually landed. It maintains state so it never repeats work on already-fixed listings, and if a run finds nothing wrong, it takes no action — a genuine no-op, not a forced update.

Critically, the agent operates inside a strict policy gate: it is only permitted to write to the description and tags fields. Any attempt to touch price, inventory, or any other field is automatically blocked and logged as an escalation for human review — it is never silently allowed. This boundary is enforced in code, not just in a prompt instruction.

How we built it

Built entirely from an Android phone — no laptop — using Google Colab as the primary development environment. The agent core (audit → generate → policy-gate → write → verify → record-state) is written in Python using the Google GenAI SDK to call Gemini (gemini-3.6-flash) and the Shopify Admin REST API for reading and writing product data. The Shopify store used for testing is a free development store seeded with a realistic mix of complete and intentionally incomplete listings, used to validate the agent's detection and fix logic end-to-end. All code, README, and architecture diagram are in the linked GitHub repository.

Challenges we ran into

Working entirely from a phone meant navigating real constraints: setting up a working dev environment without a laptop, working around Gemini's free-tier daily quota while testing, and resolving a Shopify Admin API authentication flow that required a manual OAuth token exchange after an initial wrong-token attempt. We also hit and fixed a real deduplication bug during testing — an early version of the state logic could have skipped a genuinely broken listing if it matched a previously-fixed issue type, which we caught and corrected before finalizing the logic. Separately, a Google Cloud billing verification constraint (limited access to internationally-accepted payment cards) meant Cloud Run/Scheduler deployment was still in progress at submission time; the demo shows the fully working agent running its complete detect-fix-verify loop against the live Shopify store.

Accomplishments that we're proud of

A fully working, end-to-end tested autonomous loop — including a real security boundary (policy gate) that actually blocks unauthorized writes rather than just describing one, and a verified write-then-read-back confirmation step so the agent never claims success it can't prove.

What we learned

That "agentic" is really about the decision boundary, not the amount of code — a small, well-scoped agent that reliably knows what it's allowed to do (and proves it) is more convincing than a broad one that merely generates text.

What's next for Storefront Ops Agent

Completing Cloud Run + Cloud Scheduler deployment for fully unattended scheduled runs, migrating state from local storage to Firestore, and extending the policy gate to a couple of additional safe, low-risk catalogue fields.

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Storefront Ops Agent

Built With

• ai-agents
• automation
• gemini-api
• google-cloud-run
• google-cloud-scheduler
• google-colab
• google-genai-sdk
• python
• rest-api
• shopify-admin-api