StudioLatentMatthieu Czarkowski · creative studio & engineering (home)
FREN · Passer en françaisScope a project

Radar

The platform’s editorial front: daily report, ranked ideas, production briefs

Role
Design, rebuild on the shared platform, report and briefs
Tools
  • Python
  • SQLite (socle commun)
  • HTML autonome
  • Qwen 3.5 9B (MLX 4 bits)
  • pytest
Status
In progress
The 20 October 2026 report: data freshness, then ranked ideas with their 28-day view range.
The 20 October 2026 report: data freshness, then ranked ideas with their 28-day view range.
press articles over 72 h, grouped into stories
904
ideas ranked with a P10–P90 range
16
videos in the “topics that worked” history, to calibrate the heat score
252
own collectors, databases or models: everything comes from the platform
0
The problem
Finding a rising topic and a ready-to-shoot brief every morning, without duplicating another tool’s collection and models.
For whom
The LaRefVlog creator: Radar suggests, a human decides; ideas and briefs keep the “parody fiction” label.
The outcome
An editorial front with no own collector or model: daily report, 16 ideas ranked with the Oracle predictor’s range, heat score calibrated on 252 videos, briefs and full scripts produced locally on the real data of 20 Oct.

Context

Radar is the editorial front of the shared platform: it reads the Oracle’s store and models and turns the day’s data into editorial decisions. One source of truth for both tools.

What I built

  • Standalone daily HTML report, light or dark: freshness, ranked ideas, hot topics, multi-source stories, recent videos with expected views, source health.
  • LLM-free production briefs: idea, hook, view range, why (heat, novelty), real articles behind the topic, parody label.
  • View ranges from the Oracle’s “channel regime” predictor: −21% error against the rolling median, walk-forward.
  • Heat score calibrated on a 252-video “topics that worked” history.
  • Full shot-by-shot scripts produced by Qwen 3.5 9B locally (successful job on 20 Oct), with the creator-approved style guide, through the shared GPU queue.
  • Old stack archived; 6 offline tests ensure no own collector or database comes back.

Architecture 7 blocks

  1. 01 · DataPlatformThe Oracle’s store, read-only
  2. 02 · ToolHot topicsSource agreement, freshness, velocity
  3. 03 · ModelRanked ideasP10–P50–P90 range, explained score
  4. 04 · ToolBriefsHook, why, real articles
  5. 05 · ModelScriptsLocal Qwen 3.5 9B, style guide
  6. 06 · HumanDecisionNothing is published automatically
  7. 07 · OutputReportStandalone HTML, light or dark

Pipeline 5 steps

  1. PlatformSame store as the Oracle.
  2. TopicsExplained heat over 72 h.
  3. IdeasRanked, with a view range.
  4. BriefsReady to shoot, sources cited.
  5. ReportOne local HTML page.

Decisions the choice, the reason, the alternative ruled out

  1. D1
    Remove own collectors, database and model: Radar reads the Oracle’s platform.

    One source of truth for both tools; a test checks that no parallel stack comes back.

    Ruled outSynchronising two databases.

  2. D2
    LLM-free production briefs, built from the day’s data.

    They are available even without a free GPU, and every claim points to a real article.

  3. D3
    Show the range from the Oracle’s “channel regime” predictor, not a per-idea prediction.

    It is the best predictor measured walk-forward: −21% error against the rolling median, 80% interval covered 81% of the time. The range shown is the one that holds.

  4. D4
    Calibrate the heat score on what has already worked.

    A 252-video “topics that worked” history links the signals known before publication to each video’s actual performance.

  5. D5
    No own scheduling: Radar follows the platform’s service.

    One scheduled agent for everything, installed only with consent.

Deliverables 4

  • 20 October 2026 report generated on real data.
  • Three production briefs (ranks 1 to 3) and their full scripts produced by Qwen 3.5 9B locally (successful job on 20 Oct).
  • A 252-video “topics that worked” history to calibrate the heat score.
  • Old v0 stack archived; 6 offline tests.

Environment

Hardware and measurement setup

  • Reads the Oracle’s SQLite store with the same interpreter: a single dependency stack.
  • Scripts drafted locally by Qwen 3.5 9B (4-bit MLX) through the shared GPU queue.

Reproduce README commands

Private code: read access under NDA, or a 30-minute live demo. Request access →

./vigie.sh cycle
./vigie.sh rapport
./vigie.sh scripts fiches --n 3
./vigie.sh statut
PYTHONPATH=. ../p11-oracle/.venv/bin/python -m pytest
  • Radar uses the platform’s interpreter (p11-oracle/.venv): a single dependency stack.
  • Full scripts go through the GPU queue (gpu_jobs/01).

Gallery 2

Hot topics over 72 h and multi-source stories, with their sources.
18 October version, before the move to the shared platform.

What this project means for you

The same know-how, on your problem.

01

A tailored watch

Your sources, your vocabulary, your performance data: a daily watch that scores topics and suggests, without publishing anything for you.

02

One source of truth

Your tools read the same store: the report’s numbers are the dashboard’s numbers.

03

Sourced briefs

Every idea comes with its real articles: the team checks before shooting.

Scope a project →The company offer