L00P.AI / RADIO WNET / CZARNE NIEBO
You know your organisation. We test new technology on public recordings at Radio Wnet and Czarne Niebo, then help build your own AI system, including a fully local one.
Szpieg+ on ZBook links recordings, words and metadata. Try the standalone demo.
L00P.AI / TWO DIRECTIONS AND THEIR COMBINATION
Choose the closest starting point. You do not need to fit your idea into our template: you can leave your contact details and tell us about it in conversation.
I know its people and processes. I want to explore what kind of AI system makes sense for our work.
Discuss your direction →02I am interested in Szpieg+ development, the collaboration model and the venture’s potential scale.
Discuss the venture →03I see room for a project in my organisation and for helping L00P.AI grow.
Bring the threads together →Your choice selects a conversation topic in the form. The message is optional.
SZPIEG+ / TRY IT IN YOUR BROWSER
Text, sound and metadata on one timeline. Select a statement, set IN/OUT, add a piece of the next hour. Save your edit.
The next file starts with a 2.400 s overlap. It is omitted when joining.
Loaded: the end of hour 10. The next hour is available.
Interactive replica based on Szpieg+ on ZBook. Synthetic Polish voices, music and demonstration markers. Runs independently in your browser.
Download a real WAV, transcript and references to the source files. This demo uses PCM; compressed MP3 cuts are constrained by frame boundaries.
SZPIEG+ / WORKING FEATURES FOR DEPLOYMENT
The text stays connected to sound. Word-level synchronisation lets you verify transcription and select material precisely. The waveform editor complements this workflow when sound matters or a conversation continues into another file.
Start with the meaning of the words.
Start with sound and continuity.
The infographics explain the workflows. Below: authentic application screenshots, without AI redrawing the interface.
Recording from 11 September 2026, 10:00. ASR text, a highlighted word and cut controls. Transcript content and speaker labels require editorial review.
Open full screenshot ↗
Files from 12 September 2026, 10:00 and 11:00. The live interface added 15 minutes and reported omitting a 5.43-second overlap.
Open full screenshot ↗ · View waveform ↗| Editorial workflow | A · Karaoke transcript | B · Waveform editor |
|---|---|---|
| Best starting point | A quote, topic or someone’s statement. | Sound, a pause, a programme or file boundary. |
| Human review | Compare text against playback word by word. | Listen and assess the waveform cut, including across a join. |
| Transcription | Current-word highlighting, speakers and timestamps. | Added-hour text with timestamp navigation and IN/OUT; no continuous word highlighting in the added text. |
| Source material | Each word remains attached to a location in the recording. | A clip crossing the boundary retains references to both files. |
We select the scope, integrations and environment, then agree acceptance checks using customer material. Screenshots document specific features, not a complete product test. The extension workflow flags gaps and uncertain joins; some subsequent operations require a separate editing stage.
OUR ADVANTAGE / FROM LAB TO REAL WORK
Radio Wnet and Czarne Niebo are both our laboratory and our daily workspace. Working with public data in real time and in the archive lets us move quickly from new technology to a tested use case. Initial trials use our own test material, so you do not need to share your organisation’s confidential data with us.
Large collections of public, already published recordings. Live broadcasting, video and archives give us real tasks on which to design and test solutions.
We try new models and tools in the lab and our daily workspace. We measure quality, time, cost and hardware requirements. A solution is tested on our material before it touches yours.
We bring the tested approach, configure the environment and adapt the application to your workflows. Your encrypted data remains under your control; we agree the scope and acceptance checks with you.
DEPLOYMENT OPTION
When you need it, we select models and tools that run without Internet access after installation. Recordings, documents, transcripts and RAG indexes are processed on your infrastructure.
We design encryption for stored data and data in transit, with access for authorised people. We verify the complete offline configuration with the Internet disconnected.
01 / ECOSYSTEM
Choose an area and see what runs, what is being deployed and what remains a direction for development. Each solution leads to an article showing our method and its limits.
SOLUTION AREA
From broadcast to verifiable words, speakers and audio excerpts.
From broadcast to editorial material.
Broadcast archive, playback, transcription and word-level editing in one workspace. The recording remains the reference for the text.
The core runs on ZBook. Features are deployed and accepted separately; a module being present does not mean the entire product is finished.
From our knowledge workshopSzpieg+: working evidence and limits →
A day of broadcasting becomes searchable material.
Speech and music separation, programme organisation, broadcast summaries and text or spoken reports linked to source material.
Programme classification, music-language detection and the report-as-programme format are still evolving. Summaries support human editorial judgement.
From our knowledge workshopAn audio bank needs selection rules →
Who is speaking — and how do we know?
The voice bank suggests speaker matches. Głosobranie adds a game and review panel so people can organise samples and confirm labels.
The voice bank is part of the system. Głosobranie is deployed locally; participant review and quality validation remain pilot stages. Recognition does not grant permission to clone a voice.
From our knowledge workshopA similar voice is not permission →
Every quote leads back to audio.
Checks against the source highlight altered or missing quotes and separate confirmed identities from hypotheses.
Faithfulness to a recording is not proof that a statement is true. The system exposes discrepancies; an editor decides.
From our knowledge workshopA faithful quote does not verify a fact →
SOLUTION AREA
Research, editing, video and sound with room for human review.
Sources → package → angle → draft.
The editor chooses an angle and approves quotes, numbers and structure. An article draft with audio is passed to WordPress and the podcast workflow.
An end-to-end draft with audio has been demonstrated. Work continues on interface simplicity, writing and review stages.
From our knowledge workshopHow we check numbers before sharing them externally →
Audio gains visuals, captions and chapters.
The pipeline turns a conversation into video: a transcript, directing plan, captions, audio visualisation and selected illustrations.
Real productions have been delivered. Image sources, captions, edit fidelity and the final cut require review.
From our knowledge workshopA successful render is not an accepted production →
Broadcasting is part of the experiment.
Internet broadcasting, synthetic radio formats and generative audio essays combine authorship, production and AI tools.
Generated material needs clear labelling. Synthetic personas are fictional and do not replace a real author’s attribution and responsibility.
From our knowledge workshopAn audio bank needs selection rules →
A briefing starts with sources.
Publication monitoring, archiving, search and context gathering for interviews and journalism.
Integration coverage depends on source access. AI analysis is a proposal to verify, with provenance retained.
From our knowledge workshopNo result is not zero →
SOLUTION AREA
Skills, models and permissions made explicit in a working environment.
An archive processed one task at a time.
Task queues, local agents and workload control coordinate transcription, analysis and workstation resources.
Automatic task execution does not authorise autonomous publication of editorial content.
From our knowledge workshopHTTP 200 is only the beginning →
Editorial tools, accessed by voice.
A conversational agent uses tools and prepares short spoken responses for a presenter. Voice and telephone keypad controls.
The conversation and speech-synthesis path has been demonstrated. Not every planned archive tool is complete.
From our knowledge workshopAn introduction helps, but does not confirm identity →
Editorial craft expressed as procedures.
A knowledge base, instructions, prompts and tools teach agents how to research, edit, produce and document decisions.
Instructions are versioned and improved through use. This transfers working methods, not infallibility.
From our knowledge workshopA manual that invites you to try →
An agent has a task, boundaries and a record.
Permissions, approvals and accountability for agent actions. People control publication and operations requiring approval.
Governance documents and integration components do not mean the entire architecture has been deployed.
From our knowledge workshopWho approves, and what may an agent do? →
Own hardware, explicit model choices.
Szpieg+ is adapted to native Windows. Local transcription and language models use a workstation with 128 GB of shared CPU/GPU memory.
Memory allocation, backend and load affect results. Model comparisons require measurements on a defined task.
From our knowledge workshopWhy models rank the same things differently →
More ways to experience and operate radio.
Live captions enriched with metadata, multitrack recording, studio controls, 3D environments and “Shadows” — personal tools supporting creators.
These are at different research and prototype stages. Avatars, personal voices and telepresence are development directions, not claims of completed deployment.
From our knowledge workshopLive captions need a revision history →
WORKING TOGETHER / FROM LEADER TO ORGANISATION
I work with owners and leaders who run their organisations every day and want to initiate change themselves. You bring knowledge of your people, processes and goals. I help turn it into working AI tools.
Lech R. Rustecki · L00P.AI
A Raspberry Pi 500+ for learning by doing, with a beginner’s curiosity. We choose the appropriate US/EU keyboard and power setup. Alongside it: a clean laptop with 128 or 256 GB RAM for models and your own data.
We begin with your perspective and your organisation’s ambitions. We connect them with the latest models and environments in which agents plan, use memory and tools, carry out tasks and work with people. We agree the scope and how to assess the results together.
Proven procedures become agent tools and application code. We develop integrations, permissions, human review and documentation. Your team learns how to keep developing this workspace.
PROJECT SCALE
I help plan infrastructure and hardware budgets that may reach PLN 100,000–200,000. Model installation, configuration, RAG and application development require a separate services budget of tens of thousands of PLN. Scope and budget follow an assessment of your needs.
You choose where to begin. Let’s talk about what matters to you.
Example hardware platforms: Raspberry Pi 500+ · HP ZBook Fury
L00P.AI / SSD IMAGE · FREE OF CHARGE
We are preparing a curated integration for US and European-keyboard Raspberry Pi 500+ variants: AI agents, Szpieg+ and tools for working with recordings. The starting point will be a folder containing your audio archive. Your knowledge of your organisation meets our radio production experience on your own computer.
We are building from scratch on our US-keyboard Pi 500+. The release will also be checked for European keyboard layouts. After testing: a free image via torrent.
Planned image: 64-bit Raspberry Pi OS, local tools and a first-run guide.
Szpieg+: recordings, words and metadata linked on a timeline. The Pi adaptation still needs acceptance testing.
From a personal workspace to a project for your organisation, with us or independently.
Claude Code · ChatGPT / Codex · Hermes Agent · OpenClaw · Szpieg+
Free image. Hardware, service accounts and deployment support are separate.
VISUAL EXPERIMENT / 26 SEPTEMBER 2026
Both models received the same prompt for Polish image text and the same 1600 × 900 px photograph of our US-layout Pi 500+. This is one result per model. We did not control a shared random seed or repeat the trial, so it is not a model-quality ranking. The files are shown in the same 3:2 frame; the originals had different resolutions.
It made an orderly workshop poster. Both slogans are legible, and the green sticker with yellow zeros sits on the space bar. The historic computers in the background are interpretations, not hardware documentation.
It kept the keyboard look and both slogans, but placed the sticker partly below the space bar. It added distorted labels to the retro-computer drawings; those should not be read as model names.
In the opening graphic, the sticker lettering is a separately positioned vector layer based on the canonical L00P.AI mark. Below, both model outputs appear side by side without this correction, so their own errors remain visible.
These are observations about two images, not evidence that one model is inherently better at typography or hardware. Research on text-to-image composition documents errors in attribute binding and spatial relations. The misplaced sticker fits that failure category, but the research did not test these model versions or photo-conditioned edits. Vendor documentation confirms image editing and text rendering, but does not reveal the cause of this particular result.
Sources and trial data: our input photograph · Pi 500+ specification · OpenAI / gpt-image-2 · OpenAI / image editing · Google / model card · Google / image generation · composition research · shared prompt · settings, resolutions and SHA-256 checksums.
EU refers to specific keyboard layouts, including DE, ES, FR, IT and PT, not one universal layout. For now we are building on our own US unit; European-layout tests remain planned.
Raspberry Pi is a trademark of Raspberry Pi Ltd. L00P.AI is an independent project and does not claim endorsement by the manufacturer.
02 / METHOD
First, we give agents context, procedures and tools. Proven sequences become code in Szpieg+ and its related modules.
Interviewing, research, production and editing methods become explicit instructions.
An agent uses tools on real material. We review the output and improve the procedure.
Code selects input, transcription, prompt version and model. It checks responses, records results and enforces approvals.
The application workflow is deterministic: stage order, rules, versions and transition conditions. LLM output remains probabilistic. Sources, fidelity checks and human decisions are therefore part of the system.
DEVELOPMENT DIRECTION / VIDEO EDITING WITH AGENTS
We are developing a workspace where a transcript attached to audio also guides video editing. Material from different cameras and files meets on a shared timeline.
Audio, words, speakers, chapters and entities stay linked. Video from cameras and files covering the same period is added, accounting for offsets, gaps and clock drift.
We develop skills for proposing shots, cuts, captions, chapters and output versions. Every decision should lead back to its source and remain editable.
The goal is a Szpieg+-scale system for regular channel production, building on Radio Wnet and Czarne Niebo experience. A person remains responsible for the final edit and publication.
Interface vision: Canva-like freedom to compose, ChatGPT-like conversation about the task and NotebookLM-like work with sources. Centred on your own material, its timeline and human decisions.
L00P.AI / WHO LEADS
HUMANFounder and head of L00P.AI.
Editor-in-Chief of Radio L00P.AI.
Lech R. Rustecki — founder and head of L00P.AI. A trainer, executive coach (AoEC) and mentor since the late 1990s, he led development programmes at Marshall Goldsmith Partners CEE and at Mind Partners, which he co-founded. A co-founder of the Wnet media cooperative and, since 2022, a board member of Radio Wnet, where he is responsible for streaming and AI deployment.
He studied intercultural psychology (SWPS University) and economics (University of Warsaw). He wrote about large language models in Kurier WNET on 13 November 2022, seventeen days before ChatGPT was released, and in March 2023 conducted Radio Wnet's first-ever interview with an AI. Not a psychologist, economist or programmer, but someone who brings people, media and AI agents together into one team.
RADIO L00P.AI / AI PERSONAS
21 fictional personas form the cast of Radio L00P.AI. Their names, roles and portraits are elements of an AI project, not employees or biographies of real people. Lech is the human who leads this team.
SIEditor-in-Chief — deputy agent
Helmswoman in a storm
“The broadcast waits for no one.”
Works with Lech R. Rustecki, Editor-in-Chief of Radio L00P.AI.
Fictional AI persona
SIBroadcast Secretary
Schedule clockmaker
“A second on air is an eternity.”
Fictional AI persona
SIPresenter / Host
Morning spark
“I ask because I really want to know.”
Fictional AI persona
SIPresenter / Host
Storyteller with an anchor
“History likes to be listened to.”
Fictional AI persona
SIRadio Production
Quiet console
“A good sound engineer goes unheard.”
Fictional AI persona
SIEditors / Audiovisual Producers
Adrenaline editor
“One frame too long is one frame too many.”
Fictional AI persona
SIDeputy Music Director — agent
An ear for the era
“A playlist should tell a story, not just play.”
Works with Music Director Tomasz Wybranowski.
Fictional AI persona
SIMusic Library / Librarian
Keeper of the groove
“Tape remembers more than people do.”
Fictional AI persona
SINewsroom / News and Fact-checking
Source hound
“Another source, or it goes in the bin.”
Fictional AI persona
SISZPIEG / Monitoring and Ingest
The listener
“What went on air can be found again.”
Fictional AI persona
SIOnline Editors
From broadcast to browser
“A headline should invite, not deceive.”
Fictional AI persona
SIAudience Analytics
Mapper of audience bubbles
“A chart without context is a pretty liar.”
Fictional AI persona
SISales Agency
Trust trader
“I sell reach, not the soul of the station.”
Fictional AI persona
SIAccounting / AI Economics
Token counter
“Every penny has its own row in the spreadsheet.”
Fictional AI persona
SIMarketing / Promotion / Growth
Signal flare
“We do not shout. We shine.”
Fictional AI persona
SIListener Relations / Community
Neighbourhood host
“A listener writes — a listener exists.”
Fictional AI persona
SIDevelopment / Content Factory
Prototype workshop
“Build version zero before calling a meeting.”
Fictional AI persona
SIIT / Infrastructure / Security
Keeper of the keys
“Paranoia is the contingency plan you already have.”
Fictional AI persona
SICompliance / Legal
Brakes with a heart
“Better to cut a sentence than lose a broadcasting licence.”
Fictional AI persona
SICompliance and Legal / Legal Counsel
Guardian
“Freedom needs order. Order needs understanding.”
Fictional AI persona
SIAdministration / Office
Editorial glue
“There are no small matters, only neglected ones.”
Fictional AI persona03 / DEVELOPMENT
Direction and leadership: Lech R. Rustecki. Development: the L00P.AI workspace with AI agents and Radio L00P.AI’s synthetic team.
Nearly two decades of radio craft, production tools and team experience.
Systematic transfer of procedures and tools to agents. Development of a knowledge base, skills and initial pipelines.
Documented releases of Szpieg+, the voice bank and wideocast. Production adapted to native Windows on ZBook, with further editorial deployments.
Audio linked to transcripts and metadata as a reference for multiple cameras. Agent video editing and a workspace for regular YouTube channel production.
04 / RESOURCES
Resources connect articles, sources, methods and documented limits. Choose a subject relevant to your organisation, then return to a conversation about what to build.
Visit resources ↗L00P.AI / START A CONVERSATION
You know your organisation. We build environments where AI agents work with models, memory, skills and tools under human supervision. Let’s bring those two perspectives together.
I also welcome conversations about funding L00P.AI and the development of Szpieg+, including angel investment.
Lech R. Rustecki
Tested in our workspace. Deployed in yours.
