01
AI demos beautifully and then never ships
We call a project done only when there is a working URL, database migrations, green tests and backups. Our knowledge-base assistant went from spec to production in 5 days.
The IT division of the ASTRO GROUP holding
AI agents for business: a priced register of your processes and a build-ready spec in 5–10 business days
We don't sell “AI adoption” in the abstract. We take one concrete process apart into agents with clearly bounded responsibilities and wire them into a pipeline that runs without a human. What the price covers, and what it doesn't, is written down before the work starts.
Timelines are indicative, based on our completed projects; the binding timeline, scope and price are fixed in the contract after discovery. This page is not a public offer.
Diagnosis
Why automation usually never reaches production
Over five years we've walked this road on our own systems and on client ones. Here are the four reasons projects die, and what we do about each.
02
Automation is afraid of live data — and rightly so
Bulk operations run inside a transaction with a row-count gate, a backup taken beforehand, a read-back check and a snapshot-based rollback. Money reconciliation against the marketplace dashboard closes at 0.00 ₽, order-status sync at 938 of 938.
03
The vendor leaves and you inherit a black box
On development projects, formal technical documentation and a handover package are part of the work. Transfer of the source code and of the rights to it is a matter for the contract, up to and including full assignment; the scope is fixed before we start, not after. Handover runs to its own plan: credentials, database dumps, documentation and deployment instructions.
04
One model makes a mistake and nobody catches it
Non-trivial code goes through double-blind review: two reviewers read the change independently, seeing neither the author's reasoning nor each other's findings. On critical paths one of those voices is a model from a second vendor. Blockers surface before production, not during an incident.
Directions
What we do
Six practice areas. Each one backed by a working system, not a slide.
AI assistant on your knowledge base
We build a corpus from your documentation, video and exports, then deploy an assistant that cites its sources and says “I don't know” instead of inventing an answer.
Marketplace automation
Prices, stock, listings, buyer questions, returns, reviews and ERP sync. Bulk operations without manual labour and without errors.
AI teammates for your departments
Chat operator, sales manager, bookkeeper. A function appears in the company without hiring a person for it.
Custom CRM and industry platforms
From discovery and specification to production, documentation and handover. A system shaped around your process, not the other way around.
Provable
Three numbers instead of adjectives
0days
from written spec to live production
AI assistant for an industry software vendor: a 236-document corpus, answers with citations, one deployment serving a hundred client accounts.
0/ 938
order statuses synchronised, zero mismatches
Financial-report reconciliation against the marketplace dashboard closed at 0.00 ₽ — 15 reports, 12 settlement components.
0
green automated tests on a marketplace support platform
History migrated one to one: 149,191 orders, 111,406 messages, 54,976 reviews; formal technical documentation delivered.
Practice
We deployed all of this to ourselves first
ASTROLABS grew out of the holding's own trading business. Our marketplace CRM isn't a showcase — it's the operational backbone we trade on every single day: integrations with every marketplace and with our ERP, money reconciled to the kopeck, advertising statistics where the reports used to show zeros. We aren't selling you a hypothesis; we're selling a tool we use ourselves and that has already survived contact with real data.
- Live production systems, not prototypes
- Double blind code review; on critical paths, a second-vendor model as one of the votes
- We deploy inside your perimeter on request; calls to a language model go to a provider, and exactly what is sent is fixed in the contract
- Formal technical documentation and a handover package on development projects; transfer of source code and rights is a matter for the contract
We don't walk away after launch: operations with an SLA at RUB 40,983.61 / month
A production system degrades quietly: a marketplace changes its API, the assistant starts answering worse, the backup stops restoring — and it goes unnoticed for weeks. The subscription puts your system on watch: monitoring, an incident response time fixed in the contract, updates, incident post-mortems and an agreed volume of changes per month. It costs RUB 40,983.61 per month (RUB 50,000 incl. VAT at 22%), and what is in scope — and what is not — is listed before you sign. See what the subscription covers and what it costs →
Tell us where it hurts
The first conversation is a walk-through of your process and an honest assessment: what can be automated now, what should wait, and what is better left alone.
Send a request