The adoption of AI in recruitment and HR processes is slow due to data confidentiality. Uploading sensitive candidate data or proprietary code to external clouds (such as OpenAI or Anthropic) creates real legal risks and violates GDPR. The answer to these challenges is local language models, which anyone can install on a computer with a reasonable NVIDIA graphics card. Examples include Gemma (Google), Llama, or Bielik AI — a Polish language model that we can run 100% locally.
Why Bielik AI in IT Recruitment?
Developed by the SpeakLeash team and ACK Cyfronet AGH, Bielik AI (bielik.ai) is a language model optimized for the Polish language. Unlike English-centric models, it perfectly understands Polish grammar, local technical terminology, names of Polish universities, and the specifics of the Polish IT market. Running it locally, we gain a powerful analytical tool that does not send a single byte of personal data outside the company network.
How to Install Bielik AI Locally?
Running the model on your own computer or company workstation is extremely simple thanks to the Ollama tool:
- Go to ollama.com and download the version for your operating system (Linux, macOS, or Windows).
- Install the application, then open a terminal (or PowerShell) and type the command:
ollama run bielik - Done! The Ollama API server runs in the background under
http://localhost:11434and is ready to integrate with your local ATS or Obsidian system.
What Hardware (GPU) Do You Need?
Bielik AI version 11B-v2 (e.g., in the optimal quantized version Q4_K_M) requires sufficient GPU computing power (VRAM) to generate answers smoothly:
- Minimum Requirements (GPU): A graphics card with at least 8 GB VRAM (e.g., NVIDIA GeForce RTX 3060 / RTX 4060). This allows for stable operation, although generating answers with long contexts may be slower.
- Recommended Configuration (GPU): A card with 12 GB or 16 GB VRAM (e.g., NVIDIA RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 4070, or RTX 3080). This is the ideal sweet spot between price and analysis speed (over 30 tokens/second).
- Apple Silicon Alternative: Mac computers (with M1/M2/M3 processors) with a minimum of 16 GB unified RAM. Apple's architecture handles local models excellently thanks to shared memory.
💡 Innovative Concept: NDA-Safe Code Auditor
Problem: Evaluating recruitment tasks sent by candidates or analyzing their GitHub portfolio consumes valuable hours of your Senior Developers' time. Uploading this code to cloud AI models is out of the question due to project confidentiality (NDAs) or candidate copyright.
Solution with Local Bielik: We implement a script that clones the candidate's code to a local, secure directory (sandbox). Bielik AI analyzes the project structure and automatically generates 3-5 personalized technical questions for the interview.
Example Generated Question: "In the `auth-service.ts` module, you decided to use a custom token rotation implementation instead of library X. What security risks did you identify, and how does your code protect the application against them?"
This way, the Senior Developer does not need to analyze every line of code in detail, and the recruiter receives a ready-made interview scenario checking if the candidate actually understands the submitted solution. All this without sending code to the network.
Want to implement secure, local AI solutions?
We help companies install on-premise models (Ollama, Bielik.ai), integrate them with internal ATS/CRM systems, and audit processes for GDPR compliance.
Consult with us