Scattered tools, none knew the whole year.
A master’s year is one long piece of work: lectures feed the reading, the reading feeds the thesis. My tools each held one slice, and none of them could answer a question across the rest.
Two languages
Most sport-management research is in English, my thesis is in French. Keyword search misses half of what I own.
Citations by hand
APA 7 in its French form, page numbers, a reference list that matches the text: hours of checking at the end of every chapter.
The same constraint
My thesis studies what generative AI can do with one person and almost no budget. The app is built and run under exactly that constraint.
One app, four spaces.
Thesis, library, lectures, and one question across all of them. Switch spaces below the way I do in the app, with the tabs or the arrow keys.
L’IA générative dans les associations sportives
Les associations sportives reposent largement sur le bénévolat, avec peu de salariés pour les tâches administratives (Lemaire et al., 2022, p. 52).
L’IA générative permet à un seul bénévole de produire toute la communication d’un club.
Nothing said in class gets lost.
I record the lecture while the slides play. Every slide change is timed, and I flag the moments the lecturer insists on. After class, the transcript lines up with the slides and becomes a written course.
⚑Detected in the transcript: “remember the difference between sponsorship and patronage”.
In the written course, this part ends with : one click opens the slide, ▶ replays what was said while it was on screen.
- 01Audio saved every 5 s: nothing lost if the app closes
- 02whisper.cpp transcribes on the Mac, primed with the slide titles
- 03Each sentence is pinned to the slide on screen at that moment
- 04The copilot writes the course: slides + what was said, marked “(oral)”
- 05Revision sheet and flashcards, spaced 1 → 120 days
- Flagged by me
- Detected (“remember…”, “exam”)
- Click a slide to see what the app kept
“Parrainage” finds sponsorship.
Before picking the model behind semantic search, I tested three on my own sport-management books with short French questions. Only one connected them to the English passages that answer them.
Mostly off-topic French passages
DroppedMostly off-topic French passages
DroppedThe English passages that actually discuss sponsorship
✓ Chosen · runs locallySearch combines both methods: keywords (accents, case and plurals ignored) and meaning. Results found by meaning alone are tagged as such. The model runs on my Mac with onnxruntime: indexing the whole library takes about an hour once, then two to three minutes per new book.
Useful, and still mine.
The copilot reads and never edits. Every answer is built on passages from my own documents, points to the exact page, and marks anything that is not in them.
Chapters, notes and sources are ordinary Markdown and CSL-JSON in my own folders. Obsidian, Zotero and git open them as they are.
Transcription and semantic search run on my Mac. Lecture audio never leaves it; only the few passages a question needs are sent to the model.
An hourly versioned backup in a separate history, an optional copy to iCloud or a disk, and a bin instead of deletion.
New sources, every morning.
The library does not fill itself. A second tool, the Thesis Engine, scans three research databases three times a day with a small local model, and files what matters into the vault the app reads.
- Candidate
- Kept → vault
- Escalated → DeepSeek
- Dropped
Simulation at the pipeline’s real rates: about 90 candidates a day over three passes, roughly 5% kept. The escalation frequency is illustrative.
Hover, tap or tab through any node.
- 01 Scan three databases, three times a day
- 02 Triage locally with a fine-tuned Qwen3 4B
- 03 Escalate complex analyses to DeepSeek V4-Pro
- 04 Write notes, canvases and bibliography to the vault
Nine pieces, one desk.
The thesis, tested on myself.
One person, almost no budget, and a workspace that reads, searches, transcribes and cites. A working answer to the thesis question: what generative AI can do when time and money are scarce.
- Status
- In daily use for my M2
- Built with
- Claude Code, no formal coding background
- Platform
- macOS app (Apple Silicon), or a local web page
- Fed by
- Thesis Engine, running 24/7 since October 2026
- Code
- Private repository while the thesis is in progress