Artificial intelligence
AI that works inside real systems — not just in a demo
We build AI into the products people use every day: an HR assistant that drafts orders in Tajik and Russian, document recognition, résumé scoring, speech assessment and AI sales agents. When data must stay in-house, we fine-tune and run open models on our own hardware.
- 10
- AI functions inside eKadr
- 5
- products in our portfolio with AI inside
- TJ · RU · EN
- languages our AI works in
- On-prem
- private models on our own GPU server
Flagship case · eKadr
AI across the whole HR cycle
eKadr is where our AI experience is deepest. Ten AI functions cover hiring, documents, onboarding, work planning, analytics, performance and learning — all designed around Tajik labour law and the Tajik language.
eKadr — View case studyHow Dastyor works
- 1
Employee writes a request in Tajik or Russian
- 2
The system resolves people, departments and positions from the database
- 3
A fine-tuned model prepares the order or the answer
- 4
A person checks the preview and confirms — nothing changes without it
Каримов Алишерро аз 15 октябр ба рухсатии меҳнатӣ барои 14 рӯз фиристед
Draft order prepared
- Employee
- Karimov Alisher
- Period
- 15.10 – 28.10.2026
- Days
- 14
AI assistant “Dastyor”
Turns a short request in Tajik or Russian into a ready HR order — hire, transfer, vacation, business trip, dismissal and more — or answers questions about the team. Every change is shown as a preview and applied only after a person confirms it.
Passport & ID recognition
Reads a photo or PDF of a Tajik passport or ID card and fills 16 fields of the employee card, preserving Tajik letters (Ҷ, Ҳ, Қ, Ӣ, Ғ, Ӯ). Falls back to manual entry when unsure.
Recruitment: résumés and vacancies
Scores PDF and DOCX résumés from 0 to 100, matches every vacancy requirement with evidence and suggests hire, reserve or reject. Writes vacancy descriptions in Tajik, Russian and English from a few key facts.
Work plan import
Reads an approved annual plan from a scanned PDF, Word, Excel or a photo: sections, items, deadlines and responsible staff. Matches surnames to employees and flags doubtful matches; tasks are created only after a person reviews the draft.
KPI builder
Asks up to two clarifying questions and proposes balanced KPI sets with weights, validated by the server.
Approval routes
Turns a plain-language description into an approval route built from real positions and departments.
Learning and onboarding
Generates tasks for 15 exercise types and quizzes from onboarding materials, creates voice-overs and grades written and spoken answers — with manual review as a fallback.
AI data analyst
Answers questions about the team right on the dashboard and in reports, such as turnover or headcount by department, and queries the system’s data itself.
Private AI
Our own models for sensitive data
For government and HR data we do not have to rely on public AI services. We fine-tune open-weight Qwen models with LoRA on our own examples and serve them with vLLM on a company NVIDIA DGX Spark server, behind an API gateway with per-client keys, rate limits and request logs. Each organisation can also connect its own model endpoint.
- LoRA fine-tuning on domain data (HR orders, API actions)
- Self-hosted inference with an OpenAI-compatible API
- Encrypted per-organisation keys and usage logs
- Option to run fully inside the client’s network
AI in our other products
Proven in education, government projects and sales
PMMIS — AI import with anonymisation
- Reads works-completion acts from Excel and extracts structured rows
- Replaces organisations, people, places, phones and IDs with tokens before calling the model, then restores them
- Stops the import if anything identifiable remains
VozhaOmuz — speech and learning AI
- Pronunciation assessment with Microsoft Azure Speech
- AI grading of speaking and writing with automatic provider fallback
- Dialogue trainer, curator insights with personal data stripped, AI course constructor with images and voice-over
Chatomate — AI sales agent
- Answers customers in Instagram, Facebook, WhatsApp and Telegram with a median reply time of 6.5 seconds — 104,000 AI replies since June 2026
- Understands voice messages and reads the amount, bank and date on payment receipts; works from the business’s script, files and catalogue and hands over to a human when needed
- AI tags topics in conversations and collects frequent questions for the knowledge base; usage tracking and limits per plan
- Catches forged receipts and never re-sells to customers who have already paid
Zabonomuz — AI-assisted testing
- Rubric-based grading of written and spoken answers
- Low-confidence results are routed to a teacher
- Separate consent for AI and audio processing
Our approach
Responsible AI by default
Human in the loop
AI drafts, people decide. Changes to records are applied only after confirmation.
Data minimisation
Anonymisation before external models, personal data stripped from prompts, private models for sensitive cases.
Measured quality
Evaluation sets, user feedback, usage logs and token budgets for every AI feature.
No vendor lock-in
Multi-provider routing with automatic fallback: OpenAI, Google Gemini, Azure or self-hosted models.
Built for Tajik
Prompts, data and validation are designed for Tajik and Russian, including Tajik-specific letters.
Graceful failure
If AI is unsure or unavailable, the system falls back to manual entry or human review — never to a fake answer.
Models and tools we work with
- Qwen (open-weight)
- LoRA / PEFT
- vLLM
- NVIDIA DGX Spark
- OpenAI GPT & Whisper
- Google Gemini
- Microsoft Azure Speech
- Narakeet TTS
- Embeddings
Have a process that takes too much manual work? We will help you find where AI pays off — and implement it safely.
Discuss an AI project