AI agentsscenario
Multi-agent AI systems
Not one prompt · an agent graph · calculations, documents, product scenarios
When one LLM chat is not enough
We split hard jobs into roles: one agent extracts data, another calculates, a third checks and formats the result. That is how contours like Costbl (cost of goods from drawings) and SportsAge (multi-agent match breakdown) are built.
The pain we close
- 01
One prompt is unstable on long documents and calculations
- 02
You need a transparent chain of steps, not a “black box”
- 03
Experts/formulas exist - AI should execute them, not silently replace them
- 04
You need a path from a prototype (chat) to an industrial graph
What we deploy
Decomposition into agents
Roles, inputs/outputs, tool-calling, human control on critical nodes.
Data and tools
PDF/drawings, tables, APIs, calculators, external sources.
From pilot to product
Chat/MVP first, then a durable graph, UI, billing, monitoring.
Quality metrics
Step accuracy, escalation share, calculation time, product business KPIs.
Cases
Cases for this scenario
FAVORIT
SaaS cost estimation from drawings
More than 5 months, two stages. Stage 1 - prototype: AI chat and PDF parsing (screenshots in the stage 1 block). Stage 2 - current state: full calculation (video in the stage 2 block). Product: https://costbl.ru/
SportsAge AI
AI tips for football matches
We built SportsAge AI: from a Telegram bot to a web service with leagues, match cards, AI tips, and a detailed table. Multi-agent analysis inside. Product in production: ~200,000 ₽/mo for the client for over six months. Screenshots in the gallery. Site: https://sportsageai.ru/
Project (NDA)
AI avatar for streams
We built an AI broadcast tool: on camera - a person, on air - the chosen character. Live face swap + persona select. Work demo in the “Video” block (no autoplay). Brand details under NDA.
Cost guides
Multi-agent systems are rarely “off the shelf.” We start with a pilot on one scenario, then grow the graph.
| Package | What’s included | Price | Timeline |
|---|---|---|---|
| Research pilot | 1–2 agents, accuracy check on a sample | from $2,273 | 2–6 weeks |
| Product contour | Graph, UI/API, monitoring, 2+ month iterations | by scope | from 2–5 months |
FAQ
How is this different from RAG?
RAG answers from a knowledge base. A multi-agent system does multi-step work: extracted → calculated → checked → delivered the result. RAG is often one tool inside the graph.
Is there a ready product?
For typical dialogs - ChatNeuron/CommBoost. For unique calculations and domains - a custom system, like Costbl.
See also
- →
All AI agents and AI integration
Service hub: pilot, products, stack, and FAQ
- →
On-Premise AI
Local GPU contour without the cloud
- →
AI video analytics
Computer vision for production lines
Discuss a multi-agent system
Describe the input (document/data) and the desired output - we will propose a pilot architecture.
