Live Research Demos
Real research sessions. An existing archive. Different questions and AI models.
These demonstrations show Findit6Agent working with my personal archive of roughly 35 years of journalism: tens of thousands of files, with no prior indexing, embeddings or specially prepared demo collection. Findit6Agent performs all local file operations.
Two approaches to research: In Agent Mode, capable cloud-based LLMs such as Claude, GPT and Grok can plan complex searches, decide on a strategy, and revise it as they go. In Guided Mode, Findit6Agent breaks the task into predefined steps (chunks). Smaller locally running models handle focused sub-tasks such as interpreting the question, examining aspects, evaluating retrieved documents and synthesizing an answer. The workflow is defined in the program rather than planned autonomously by the model.
All demonstrations use the same real archive, accumulated over 35 years of work as a German television journalist. Naturally, the questions were chosen because this archive contains relevant information. These are examples, not a specially prepared dataset: a historian, researcher, lawyer or business owner might ask entirely different questions about their own files. The subject changes, but the principle is the same. Results depend on the material, the model and the question 鈥?imagine what you could investigate in your own archive.
The questions were prepared, but the research runs are genuine. Some waiting periods in the video have been accelerated and marked with >>. The published protocol shows the original processing timestamps and the AI's result, including any errors.
01 路 Axel Friedrich: Career, Collaboration and Wood Stoves
In a live session, Claude Sonnet researches the career of German environmental expert Axel Friedrich, his long-standing professional connection with journalist Michael Houben, and an unexpected third topic: emissions from wood-burning stoves.
The local full-text search finds 224 matches in under a minute. Sonnet then receives excerpts, assesses the candidate sources, reads three documents in full and produces a report citing specific archive files.
02 路 A Phone Number Hidden Among 50,000 Files
Can an AI locate one specific detail buried in a tiny text file inside decades of accumulated documents? Perplexity searches Michael Houben's unchanged 54.7 GB journalism archive for one person's mobile phone number.
Its first search mixes NEAR and AND without the necessary parentheses. Findit6 reports the ambiguous expression; Perplexity fixes the query on its own. The corrected full-text search yields 13 results. The agent opens a single TXT file and identifies the requested number, with a source reference.
The mobile number is masked in the published video and transcript.
03 路 SF6 in Wind Turbines: Investigating a Climate Controversy
What is the highly potent greenhouse gas used in wind turbines, and how did the political controversy around it develop? GPT Sol searches a journalist's original archive of reports, scientific studies, industry statements, EU policy discussions, correspondence and audience reactions.
GPT Sol carries out two full-archive searches, broadening the second one to cover F-gas regulation and political proposals. It reviews excerpts, reads selected original sources and writes a detailed, referenced report about SF6 emissions, technical alternatives and the debate about restrictions. The report explicitly distinguishes what the historical archive can establish from current legal developments it cannot verify.
04 路 Counting a Decade of TV Collaboration: Grok Keeps Digging
How often has journalist Michael Houben worked with editor Alexander Kraus (also credited as Krause), since when, and in what role? Grok investigates a decade of productions across a 35-year archive, attempting to distinguish real collaborations from name collisions, drafts, scheduling requests and alternate edits of the same report.
Five full-archive searches, repeated excerpt analysis and targeted checks of misspellings and production identifiers lead to an itemized list. Grok even notices an inconsistency in its own yearly totals and revises its initial count from 60 to 59 independently documented productions, with explicit borderline cases. It identifies one camera credit among predominantly editing assignments.
The most recent collaboration was not yet stored in the searched archive. The video is a narrated, time-compressed presentation of the original research session. Preview image is a placeholder.
05 路 SF6 in Wind Turbines: Local Qwen Builds a Search Query
How can a small local AI model find a substance it has not identified yet? Unlike powerful cloud models that can independently plan and pursue complex research, Guided Mode gives smaller models a predefined sequence of manageable steps. This demonstration makes those steps visible: the question is separated into aspects, translated into AND/OR conditions, optimized with truncation and expanded with alternative terms before Findit6 searches the local archive.
The question revisits the SF6 topic investigated by GPT Sol in video 03. Here, Qwen3 14B running through Ollama starts without naming SF6 in the search expression, then learns the answer from documents found by Findit6. The final response is shorter and less detailed than the premium-model report 鈥?but this entire workflow runs locally on an RTX 4060 Ti with 16 GB of VRAM.
The original research session and the model鈥檚 answer are reproduced in the protocol, including their imperfections. Waiting periods in the video are visibly accelerated rather than omitted.
More demonstrations coming soon
Additional research runs with different questions and AI models may be added here. All use the same long-established personal archive rather than purpose-built sample data.