AI automation for manufacturing and agriculture

Cut operating costs and grow margin

We take one business process end to end and put it into production in 1–2 months. Your team stops doing it by hand.

  • Manufacturing, 130 sales reps
  • AgroTech, 1,000+ head of cattle
  • Real estate, 15,000 listings

Where Cortex IT runs in production

Woodworking and metalworking machinery distributor. 130 reps, sales across Russia. Live since 05.2026.

Dairy farm, 1,000+ head. Computer vision watches for key events across 4 cameras, infrared at night. Live since 03.2026.

Service for real-estate agents, ~15,000 listings. Telegram bot: plain-language search, client profiles, objection handling. Live since 09.2025.

Python · RAG · Qdrant · YOLOv11 · Bitrix24 · 1C

Cortex IT by the numbers

0

sales reps using it in production at one client (manufacturing)

0

documents in a RAG index with hybrid search

0

projects delivered

0

years in the IT industry

Live in production

What already works for clients

Cortex IT case: machinery distributor

Woodworking and metalworking machinery distributor. 130 reps, sales across Russia.

situation

The knowledge base held 11,000 PDF/DOCX/PPTX files and 60 training videos — around 120 GB. Looking up the specs of one machine took a rep several minutes.

task

Answer any question about any machine model in seconds, inside Bitrix24. Everything stays on servers in Russia.

action

A RAG assistant inside the Bitrix24 messenger: Python and FastAPI, Qdrant with hybrid search. Semantic plus BM25 — that combination matters for model numbers like STD-120. Celery handles the queues, all of it in Docker on a Russian VPS.

result

Reps get answers in seconds instead of minutes, in the tool they already use. Onboarding a new rep got shorter too — the assistant answers what a colleague used to.

  • Python
  • FastAPI
  • Gemini 3 Flash
  • Qdrant
  • Bitrix24
  • Docker
  • Celery

Delivered: May 2026

Read the case

Cortex IT case: dairy farm — animal identification

Dairy farm, 1,000+ head. Photo identification plus Re-ID from video.

situation

Staff identified animals by hand, reading ear tags. Over 1,000 head, and every cow costs time. Health parameters and treatment records lived in separate systems.

task

Point a phone at a cow, get her record — parameters and treatments in one card. Re-ID from the video feed with nobody watching the screen.

action

EfficientNet-B1 with 128-dimensional embeddings, plus ByteTrack. From a photo it identifies the animal at 78% accuracy and returns her parameters and treatment list. From video it runs Re-ID across 4 cameras in real time. Three interfaces: mobile app, web panel, API.

result

78% accuracy from a photo, and the animal's record opens in one tap. Re-ID runs on the video feed without a human watching. Shipped in 2025.

  • Python 3.12
  • PyTorch
  • EfficientNet-B1
  • ByteTrack
  • OpenCV
  • FastAPI

Delivered: November 2025

Read the case

Cortex IT case: dairy farm — event monitoring

Dairy farm, 1,000+ head. 4-camera system, infrared event monitoring at night.

situation

Catching key events meant someone on watch around the clock. Night shifts, fatigue, missed events. Every miss costs money.

task

Detect key behavioural events automatically and warn the herd manager hours before they happen.

action

Dataset pipeline: video → frames → annotation with CLIP and Roboflow. Detection runs YOLOv11 plus RTMPose (17 pose keypoints) and a BiLSTM with attention for temporal patterns. Four cameras, bird's-eye projection, infrared mode after dark. Deployed on a Jetson Orin NX at the edge — no cloud.

result

Warnings land 1-9 hours ahead of the event, with target accuracy of 83-100%. The system runs on the farm itself, so a dropped connection changes nothing.

  • Python 3.12
  • PyTorch
  • YOLOv11
  • RTMPose
  • OpenCV
  • Ultralytics
  • Jetson Orin NX

Delivered: March 2026

Read the case

Cortex IT case: real-estate agency

Service for brokers and developers, ~15,000 listings, Novosibirsk.

situation

Agents searched developer websites by hand — minutes per request. The client cools off while they search. Nothing about the client got saved anywhere.

task

Let agents search in plain language, keep the listings database synced across every developer in real time, and give them client profiles and sales arguments in the same place.

action

Phase 1 was an MVP on n8n and Supabase: an LLM pulls parameters out of free text → SQL → listing cards. Phase 2 moved to a Python backend: 7 entities from the Realty API, a sweep function, client lookup by phone number, and objection handling in three tones.

result

Search takes seconds instead of minutes. The agent gets ready-made objection responses. The bot doubles as a CRM without leaving the chat.

  • Python
  • Telegram Bot API
  • LLM
  • PostgreSQL
  • Supabase
  • n8n
  • pytest

Delivered: September 2025

Read the case

How we work

Four steps from first call to sign-off

Intro call (30 minutes)

What happens

questions about your process, no pitch. Nothing to prepare.

What you get

a clear view of which process AI closes fastest.

Mini-audit (2-3 days, free)

What happens

a two-page written document: what we saw, what we propose, a link to a similar case.

What you get

a written position you can take to your owner or your board.

Proposal (3 options)

What happens

Base / Recommended / Premium. Valid 14 days. We walk you through it on a call.

What you get

budget and timeline broken down line by line.

Contract, 50% upfront, milestones with sign-off

What happens

signed through Saby or Diadoc. No work starts before the deposit. Each milestone closes with an acceptance act.

What you get

a project with defined milestones, paid as they are delivered.

FAQ

Questions we get most often

Next step

Request a mini-audit — 2-3 days

A two-page written document: what we saw in your process, and what we propose. Free.

  1. We read your form within 4 hours and reply on Telegram or by email.
  2. We book a 30-minute intro call — questions about your process, no pitch. Nothing to prepare.
  3. Within 2-3 days you get the written mini-audit, with a link to a similar case.

20-2000 characters

Telegram, email or phone — we reply within 4 hours