Zapyt

Challenge

Every day people write in public: "can anyone recommend an English school for adults?", "who has booked a tour to Egypt through an agency in Kyiv?". Each of those posts is a warm lead, and almost every business misses them:

• They are scattered. Hundreds of Telegram chats and channels, Threads and X, with no single place where intent shows up.
• They go cold in hours. The first useful answer under a post usually wins. Checking feeds by hand once a day is already too late.
• A template reply kills it. People see "write to us in DM" from five competitors under the same post and ignore all five.
• Automation is risky. Mass commenting gets accounts banned and burns the brand.

We built Zapyt as DForce's own product to solve this, and run it for clients in education and travel.

What we built

A multi-tenant SaaS that works as a lead radar plus a copywriter, with a human approving every message:

• Collectors on three platforms. A Telegram collector on a user account monitors chats in real time and runs keyword search. Threads and X collectors work through the client's own account in a headless browser, with per-client proxies, a human rhythm, run slots and a daily budget, so the account looks and behaves like a person.
• Qualification by an AI agent per client. Every campaign has its own agent spec: who the ideal customer is, domain knowledge, guardrails and examples. Claude reads each post, decides whether it is real intent, scores it and explains why.
• A reply in the client's voice. For each lead the system drafts a specific, non-template comment and private message that answer the actual question. When a post already has many replies, it is marked as contested and the draft has to stand out.
• Approval in one tap. Leads arrive as Telegram cards with Approve / Edit / Skip, and in a web cabinet with filters, a "hot" sort and analytics.
• Results are tracked automatically. The collector goes back to the post, checks whether the person answered our comment and counts likes, so the funnel is measured, not guessed.
• Learning from edits. When the client rewrites a draft, the edit becomes an example for the agent, and drafts get closer to the client's own style.
• Digests and reports. A morning digest at 8:30 with the top leads of the night, and a weekly report on Monday: found, sent, answered, enquiries.
• Self-service account connection. Clients log in to Threads and X right from the cabinet through a streamed browser view. The password never reaches the operator.

What the service deliberately does not do: no mass spam, no scraping of group member lists, no action without a human approving it.

Results in numbers

Measured in the production database, July to September 2026. Clients are shown by niche, without names.

• 155,400 posts analysed and 3,617 leads found, 1,912 of them in the last 30 days.
• 6 out of 7 drafts are sent without edits (746 of 871, 85.6%). The client spends seconds per lead, not minutes.
• 14% of people answer the comment (63 of 442 leads with a tracked outcome).
• Online English school: 27 leads in the first 3 days, 999 in total, 46 replies out of 367 tracked and 2 enquiries or sales. Best month, August 2026: 235 comments sent, 42 replies, 2 enquiries.
• Kyiv travel agency: 38 leads in the first 3 days, 22 replies sent, 4 of 4 tracked replied.
• English school for adults: 93 leads in 3 days, 155 in 7 days.
• Travel service for Ukrainians: 24 leads in the first 3 days, 799 in total.

How we worked

The first commit was on 24 June 2026, and the first production deploy followed five days later. From there the product grew in short releases, each one tested on live client campaigns: Telegram first, then Threads, then X, then outcome tracking, learning from edits and the cabinet login.

The principles behind it:

• No number without data. Any claim on the product site is recalculated from the database before release. A case without tracked replies is not published.
• A human stays in the loop. The AI finds, scores and writes. A person decides what gets posted.
• Tests are the safety net for fast iteration. Thousands of automated tests across the edge functions, collectors and cabinet, and over a hundred database migrations in under four months.

Stack: Supabase (PostgreSQL, Auth, row-level security, pg_cron), Deno Edge Functions (TypeScript), Next.js, Python (Telethon), Node.js (Playwright), Claude, Telegram Bot API, Netlify.

Service: AI product development