Applied AI research at brillQ
At brillQ, an AI orchestration company in Warsaw, we follow AI research closely and put what holds up into practice. Our focus is multi-agent orchestration: how teams of AI agents can be orchestrated, evaluated and trusted with real work.
We don't do R&D to publish papers. We do it to build a real advantage for our clients.
Research areas
Agentic AI Systems
We build systems of AI agents that plan, reason and carry out multi-step tasks with tools, with human checkpoints where they matter. This is the core of how brillQ works.
- Multi-agent orchestration
- Planning and task decomposition
- Tool-using agents
- Human review and verification
Large Language Models
Language models are the reasoning engine inside every agent. We build on Claude and on open models such as Llama and Mistral, and study how to choose, adapt and ground them for a given task.
- Model selection and evaluation
- Fine-tuning with LoRA
- Retrieval-augmented generation (RAG)
- Prompt and context design
Vision & Multimodal AI
Real work involves documents, diagrams and images, not only text. We research vision and multimodal models that let agents read and act on visual input.
- Document and diagram understanding
- Object detection
- Semantic segmentation
- Vision transformers
MLOps & Infrastructure
Research into running AI systems reliably in production: monitoring models and agents, evaluating changes, controlling cost and deploying where the data lives.
- Model and agent monitoring
- Evaluation and A/B testing
- Cost and latency control
- Edge deployment
These areas are the research side of the AI competences we work in for clients.
R&D in the Liquid Enterprise model
Liquid Enterprise is how we run brillQ: a small core of people sets the direction, and teams of AI agents assembled per task do the execution. In R&D, that means no slow, sequential research process. Experimental swarms test hypotheses in parallel.
Every hypothesis is checked by an architect before it moves forward, and what holds up is reused as Core IP instead of rebuilt. Research feeds straight back into how brillQ works.
- 01
Experimental swarms
Parallel explorationExploring new architectures in parallel
Dedicated research swarms test different AI architectures, algorithms and approaches side by side. Each swarm works on its own hypothesis.
- 02
Fast validation
Human reviewHypotheses are tested quickly and checked by people
No approval committees and no long queues. Swarms test feasibility and analyze the results, and our architects review the evidence before anything moves forward.
- 03
Core IP synthesis
Orchestration layerPatterns that work become Core IP
Proven solutions become reusable components of our orchestration layer: agent configurations, workflows and review steps we can apply again instead of rebuilding them.
- 04
Client deployment
ReuseNew capabilities go into client work
New Core IP flows into our client deployments and into the orchestration layer we use in every engagement. See how an engagement runs.
- 05
Continuous improvement
Feedback loopsEvery deployment teaches us something
Each deployment produces feedback. Swarms analyze performance and surface bottlenecks, and our architects decide what changes in the Core IP.
Traditional R&D is a linear waterfall: exploration → validation → prototype → product → deployment.
Our R&D runs on parallel swarms that experiment, get reviewed by people and turn what works into Core IP we use in client work.
What our R&D means for clients
Each engagement starts from what our research has already tested, not from scratch. Three things flow from R&D into client deployments: the models, the methods and the operations described below. Feedback from those deployments decides what changes in the Core IP next.
Models
The models we build on: Claude, and open models such as Llama and Mistral, chosen and adapted for the task.
Methods
The retrieval, fine-tuning and evaluation methods we test before they are used in an engagement.
Operations
The monitoring and cost controls we research, so systems keep running reliably in production.
This is the research side of the Liquid Enterprise model: the same architects who review the research design the AI orchestration services we deliver with it. The terms we use are defined in the glossary.