The problem

A D2C jewelry brand was pushing hard on a new B2B reseller program — bulk pricing, dropship, private-label options. Interest exploded, but so did the volume of questions: MOQs, lead times, catalog access, credit terms, sample availability. Their small sales team was drowning in first-touch responses, and prospects who didn't get a fast reply simply ghosted.

They needed a Telegram bot that could handle first-touch qualification 24/7, answer 80% of product questions from a live catalog, and hand only warm leads to sales — with enough personality to feel like a real brand experience, not a chatbot.

The solution we shipped

We built a Telegram AI sales assistant powered by Google Gemini Flash, backed by the brand's product catalog and reseller program docs (retrieval-augmented generation). It answers product, pricing, MOQ, shipping, and terms questions in the brand's voice, then qualifies the prospect through a light conversational flow (business type, expected volume, geography, urgency) before offering to route to a human sales rep.

To give the bot warmth — critical for a jewelry brand where feel matters — we added voice replies. Users can tap "listen" on any bot response and get an on-the-fly voice message generated with edge-tts, mixed and normalized with ffmpeg, and sent as a native Telegram voice note. It transformed the bot from a wall of text into a conversational experience.

Key features

  • Gemini Flash-powered conversation with brand-voice system prompt
  • RAG over the live product catalog and reseller program documents
  • On-demand voice replies (edge-tts → ffmpeg → Telegram voice note)
  • Multilingual: English & Hindi at launch, easily extended
  • Lead qualification flow with structured capture (business type, volume, region, urgency)
  • Handoff to human sales rep with full conversation transcript
  • Admin controls: catalog refresh, banned-topic list, response tuning
  • Analytics: intent breakdown, drop-off points, qualified-lead count

Stack

  • LLM: Google Gemini Flash (cost-effective, fast enough for chat)
  • Voice: edge-tts for neural voice synthesis, ffmpeg for opus/ogg conversion Telegram expects
  • Framework: Python 3.11 with aiogram (async Telegram bot framework)
  • Retrieval: Lightweight vector index over catalog + docs (updated on catalog change)
  • Storage: PostgreSQL for conversations, leads, and admin state
  • Hosting: VPS with systemd, autoscaling worker pool for voice generation

Result

First-touch response time dropped from hours (during business time) or overnight (out of hours) to a few seconds, around the clock. The sales team went from answering the same MOQ / lead-time / catalog questions dozens of times per day to receiving pre-qualified leads with a structured summary. Voice replies became an unexpected engagement driver — prospects would send voice notes back, giving the sales rep useful context (region, tone, urgency) they'd never have gotten from text alone.

What made this project work

Not treating "AI" as a bolt-on. The bot's system prompt was written iteratively with the client's sales lead — refined until the answers actually sounded like the brand. RAG kept the bot honest (grounded in the real catalog, not hallucinated). And voice output was the differentiator — technically simple to add, but transformative for a category where warmth matters. The whole project shipped in about three weeks.

Want an AI sales bot like this?

AI Telegram bots (LLM + RAG + optional voice) start at $799, delivered in 2–3 weeks. Includes GPT / Claude / Gemini integration, custom knowledge base, and full source code.

Get a Quote

Related: AI Telegram Bots service · AI Chatbot Guide · E-commerce Automation case study