Multilingual inbound across eleven languages. Warm-up discovery, twenty-four hour follow-up, and a wallet-scrub layer that strips crypto-scam bait before a customer ever sees it.
7,128 messages · 883 customers · 17 days
Production LLM agents that work inbound on Telegram, WhatsApp, Instagram and Discord, and wake a human only when the deal earns it.
Real outbound, anonymized. Channel, language, reply latency.
One agent, seventeen days
It does not open with a menu. It does not say it is an assistant. It answers the question that was asked, in the language it was asked in, at the length a person would use.
Language detection runs on a fastText classifier, not a guess. Reply timing is randomised inside human bounds. When a conversation stops being a support ticket and starts being a deal, it hands you the thread with the context attached.
Cumulative, all deployments
Totals since launch. The dashboard behind this link shows the same figures live, per day and per channel.
Running now
Multilingual inbound across eleven languages. Warm-up discovery, twenty-four hour follow-up, and a wallet-scrub layer that strips crypto-scam bait before a customer ever sees it.
7,128 messages · 883 customers · 17 days
On the official Graph API rather than a scraper, which is why it still runs. Multi-tenant OAuth, cleared Meta App Review and Business Verification.
Official API · Multi-tenant
Sells broker partnerships in a regulated vertical, so it never quotes terms in chat. A backstop classifier catches meetings the main path misses, which it has done.
3 meetings in 11 days · 0 crashes
Six model providers behind one abstraction with automatic failover and a per-call cost ledger. Hybrid retrieval on pgvector and SQLite FTS5, answering with the source attached.
47 cron jobs · Self-provisioned VPS
Most of the work is the plumbing behind it. Scraping and enrichment, CRM writes, document pipelines that turn unstructured reporting into queryable data, scheduled jobs that reconcile things nobody wants to reconcile by hand, and voice in and out where a channel needs it.
Python, FastAPI, PostgreSQL, Redis, Docker, Linux, n8n. Claude through the Messages API, the Agent SDK and MCP, alongside OpenAI, Gemini, DeepSeek and local models on Ollama when data cannot leave the building.
I also run five trading systems of my own against MT5, Interactive Brokers, Bybit and Alpaca. That is where the paranoia about silent failure comes from.
What things cost
| Engagement | Scope | From |
|---|---|---|
| AI sales agent | One channel, your tone, your rules, escalation to you. Runs and is maintained.Additional channels priced per channel. | $300 / month |
| Landing page | One screen, built and deployed. Your copy.Turnaround measured in days. | $499 |
| Full site | Design, copy, build, deploy, and the domain wired up.End to end. No handoff to a third party. | $1,500 |
| App MVP | A working product, not a prototype. Backend, data, auth, deploy.Scoped before anything is charged. | $6,000 |
Consulting and longer engagements bill at $90 to $125 per hour. If a fixed scope is cleaner for you, say what you need and I will quote it as one number.
Then read the transcripts and decide. That is a faster answer than any call I could give you.