SaleBuy

Client-Side Image Compression

Camera photos are 3-5MB. Sending raw files to Gemini wastes bandwidth and increases response time. We compress images client-side using HTML5 Canvas resampling — reducing file size to under 1MB without visible quality loss. This cut API latency by 50% and makes the feature usable on mobile data connections.

Structured Prompt Engineering

The Gemini prompt explicitly requests raw JSON with specific field names (title, description, category, condition) — no markdown, no commentary. We tested 15+ prompt variations to find the format that produces consistent, parseable output. The server-side parser has fallbacks for the 10% of cases where the model wraps output in code fences or adds explanatory text.

Form Pre-Population Flow

The parsed JSON is immediately injected into the React form state. The seller sees a completed draft in under 10 seconds — title, description, category all filled in. They review, adjust if needed, set a price, and post. The cognitive load drops from 'write a listing from scratch' to 'verify and price.'

Seller-Buyer Real-Time Chat

Once a listing is live, interested buyers can open a chat thread directly on the product page. Messages are stored in MongoDB with user references. The chat interface shows online status and supports multiple concurrent conversations per listing.

Parsing Non-Deterministic AI Output

LLM outputs aren't consistent. Gemini sometimes wraps JSON in markdown code fences, sometimes adds explanatory text before the JSON, and occasionally uses different field names than requested.

Solution: A multi-layer parser: (1) strip markdown code fences with regex, (2) find the first valid JSON object using bracket matching, (3) normalize field names to our schema, (4) apply defaults for any missing fields. The user always sees a complete form — the parser handles the AI's inconsistencies silently.

AI Endpoint Cost Protection

Each Gemini API call costs money. A bot creating thousands of fake listings could drain the API budget in hours.

Solution: Three protection layers: (1) Next-Auth session requirement — you must be logged in to create listings, (2) per-IP rate limiting — max 20 AI calls per hour, (3) file type validation — only JPEG/PNG/WebP are accepted. Combined, these prevent both automated abuse and accidental over-use.