LeadFlow AI — commerce lead & order agent

Small online shops in Central Asia sell through chat. Owners drown in mixed-language messages and lose hot buyers while answering price questions. LeadFlow AI is an agent that sits between the shop's channels and its owner. Every incoming message goes through an open LLM served on AMD GPUs through the AMD Token Factory API (OpenAI-compatible, DeepSeek-V4-Flash by default — one API key switches the app live). The same client can also be pointed at a self-hosted vLLM server on an AMD Instinct MI300X with ROCm for dedicated deployments. The agent classifies intent (order, quote, question, complaint, spam), extracts products, quantities, budget and urgency, writes a reply in the customer's own language, and when the customer wants to buy, creates an order draft automatically. Leads are ranked by a hybrid score: deterministic rules (contact quality, channel, keywords, de-duplication) plus model-extracted features, with a human-readable reason for every point. Model output is strictly validated, and if the LLM endpoint fails the system degrades to an offline analyzer instead of going down. An n8n workflow connects a Telegram shop bot: customer writes → agent analyzes on AMD → reply sent → owner gets a ping for hot leads. A dashboard shows ranked leads, order drafts and live inference metrics (latency, tokens/s). Because the MI300X has 192 GB of memory, one GPU can serve many shops concurrently, or a much larger model without sharding — a cheap path to a multi-tenant SaaS for SMB commerce.