We build agents, RAG and automations that earn money and run in production — not ones that look pretty on slides. We join your team or take the whole project — from architecture to delivery.
Autonomous agents in a loop doing real work — analysis, decisions, automation — in production, not in a demo.
Retrieval over your documents and data; grounded answers, auditable, backed by evidence quotes.
LLM-powered document and process automation — with explainability: you know why the model made each call.
Training and programming robots with agentic loops — embodied AI that does hard work, not a show.
We join your team for a concrete project — when hiring can't keep up; we plug into your stack and process and work as your people.
We take the project from architecture to production and deliver the whole thing — you get the outcome, not a hiring problem.
Behind every delivery stands a tight task force of experienced freelance specialists who have shipped together for years — ML, backend, frontend, robotics, RAG, DevOps. Szymon runs it hands-on and matches the right seniors to each problem.
Ex-CTO of uPacjenta.pl (with an exit), co-founder of Diagnostyka Digital Hub, founder of RIOT — the first Polish Digital Product Studio — and former Head of Product at Golem Network.
You pay for work actually done — no more, no less; flexible scope. Best when the scope is still moving.
Agreed scope, agreed price; budget certainty. Best when the scope is well defined.
The brand needs to know whether the messages appearing online are consistent with its PR strategy — scoring hundreds of publications by hand is slow and inconsistent.
An agent pulls publication links from media monitoring, extracts article content and metadata, then scores it against the brand's PR and brand theses (a brand playbook auto-built from the brand's documents as a RAG).
Auditable verdicts: aligned / neutral / divergent — each backed by an evidence quote (proof of concept).
Investigators need to surface evidence and connections from phone data — including hours of audio recordings to review.
A React/TS desktop app plus an audio transcription pipeline: cloud models via Replicate, speaker diarization, VAD chunking, and LLM correction of grammar and speaker labels.
A working application: surfaces evidence and connections from phone data along with analysis-ready transcripts.
Every supplier sends a tender Excel file with a different header format — normalizing them by hand is slow.
The system detects tables in Excel, infers canonical tender headers with GPT, maps them structurally, and generates normalized, styled spreadsheets plus Markdown explainability logs (source→target mapping, confidence %, and the model's reasoning when confidence is low).
Runs in production: normalized spreadsheets plus explainability logs for every decision.
Physical security needs autonomous, always-on robotic monitoring — without a person constantly on site.
Agentic AI drives a robotic platform in a perception → decision → action loop, combining machine vision with autonomous decision-making.
A built autonomous robotic security system running in a perception → decision → action loop.
Tracking-grade spatial analysis (pitch control, pressing, off-ball-run value, line-breaking) is out of reach for CEE and lower-tier clubs and agencies — they only have cheap on-ball event data.
Reconstructs the off-ball spatial layer from cheap event data: a physics-based spatial model (positions → pitch control etc., no training data needed) plus a light learned imputation step, calibrated and validated against paired event + tracking data.
The make-or-break technical gate passed: a player-level Spearman > 0.7 target, achieving 0.83 (LOMO) on sealed paired matches under a pre-registered, adversarially-audited protocol.
Training and programming a humanoid for an international robotics tournament — within a tight time window.
The entire robot training and programming was run by AI agents in a loop — the agents iterated on the robot's behavior right up to the start.
3rd in Poland, top 32 of 75 teams from around the world — training fully run by agents in a loop.
We understand the problem and the business context — what should happen and how we'll know it worked. You get a clear scope and goal.
We design the solution: choice of models, data and integrations. You get the architecture and a plan before we start building.
We build a working PoC or feature and iterate on real data. You get something that works and that you can actually touch.
We ship to production and hand the project over with docs plus maintenance. You get a living system, not a demo.
A short call, a straight answer: whether it's doable, how, and in which model. One contact, one accountability.