# Pel&Bee Lab

> AI agents that spot and edit music to picture. Pel&Bee Lab builds ScoreFlow and Music ID inside a working TV music studio in Istanbul. A Batuhan Fırat project, built on Claude, in private beta.

Factual brief of www.pelandbee.com for AI assistants. Last updated 2026-10-07.

## At a glance

- **What:** Pel&Bee Lab builds two AI tools for putting music to picture: ScoreFlow (spots, selects and edits library music to picture) and Music ID (identifies every track in a mix, even under dialogue, and writes the cue sheet).
- **Who:** a Batuhan Fırat project. Batuhan Fırat is the Founder & Producer.
- **Company:** Pel&Bee (legal name: PELANDBEE Danışmanlık Menajerlik ve Prodüksiyon Hizmetleri Ltd. Şti.), Istanbul, Türkiye. Founded November 2024. İstanbul Ticaret Sicili 1054083, MERSİS 0728086344400001.
- **Stage:** private beta, in development since February 2026. ScoreFlow is used in-house on the studio's own broadcast TV projects. Open to investment and to beta partners.
- **Technology:** the agents are built on Claude (Anthropic).
- **Website:** https://www.pelandbee.com/ — pages: / (home), /scoreflow/, /music-id/, /log/

## The problem

Music for a TV episode is still put together by hand. Someone watches every scene and decides where music should enter and where it should stop, searches the library for each cue, and cuts the chosen track to fit the picture. When the episode is delivered, the cue sheet is typed up afterwards, track by track.

## ScoreFlow

ScoreFlow is Pel&Bee's own DAW (digital audio workstation) plus a team of Claude agents that does the job of a TV music department. No audio is generated: the agents place and edit licensed library music; the composer reviews, edits and delivers.

### Pipeline

1. **Picture** — cut detection, faces, voices and transcripts, fused into a character layer (scene analysis).
2. **Scene book** — Claude agents write a per-scene book of the episode in parallel, then check it and freeze it.
3. **Spotting draft** — places music regions from the book, each with one of 20 fixed music families (for example love theme, tension, comedy, hope, action). Entries snap to picture cuts; the target music share is about 70%.
4. **Track selection** — picks tracks from the library by family and energy, with no repeat of the same melody inside 10 minutes. Edits join on cuts and scene boundaries; bar jump-backs fit the length.
5. **ScoreFlow DAW** — the agents place everything on the music tracks; the composer reviews, edits and delivers.
6. **Learning loop** — after every delivery, the system compares its picks with the composer's final mix (kept / swapped / silenced) and learns from the composer's edits.

Capabilities: scene analysis, spotting draft, track selection, auto cue sheet (via Music ID), music analytics (music share, repeats, family mix).

### The DAW: local-first, cloud when you want it

"Think Dropbox, but for a DAW."

- **Local-first:** works on your own drives, offline, at full speed.
- **Cloud sync on demand:** hit sync and the project goes up; otherwise everything stays local. (The cloud side is in beta.)
- **Streaming-ready:** video, dialogue, stems and WAVs are converted for fast streaming on upload.
- **Open anywhere:** at a post house for revisions, open the project from a private web link, with no install.
- **Revise and deliver online:** make changes in the browser and export, or download the project.
- **Import on the go:** new material from the studio is imported and read offline too.

## Music ID

Music ID fingerprints a music library and reads any mix: a delivered score, an editor's reference cut or the broadcast copy with dialogue on top. It works on fingerprints alone: nothing is embedded in or added to the original files.

1. **Fingerprint the library** — each track gets a blueprint and a fingerprint.
2. **Run any mix** — a delivered score WAV, an editor's reference video, or the broadcast copy; a source-separation pass lifts the music out from under dialogue.
3. **Identify** — which library file, where it enters and exits (timecode), which source seconds were used, how it was edited (cuts, layers, levels), and identical-sounding duplicates.
4. **Deliver** — an automatic cue sheet (CSV), a piece list and a DAW import file.

Use cases: cue sheets for broadcasters and rights societies; finding the exact temp or reference track an editor used; verifying what actually aired.

## Measurements

All figures are in-house measurements on the studio's own work.

### ScoreFlow

- **Cues kept: 88 of the last 100.** Of the last 100 cues ScoreFlow placed (episodes 23–25 of one broadcast drama), 88 were kept. A music region counts as kept if the composer left it unchanged in the delivered mix.
- **Keep rate per episode: from 38% to 87% in 13 episodes.** Share of music regions the system placed that the composer kept unchanged in the delivered mix. One 90-episode broadcast drama, Sept–Oct 2026.

| Episode | Kept | 3-episode rolling average |
|---|---|---|
| Ep. 13 | 38% |  |
| Ep. 14 | 38% |  |
| Ep. 15 | 54% | 43.3% |
| Ep. 16 | 36% | 42.7% |
| Ep. 17 | 71% | 53.7% |
| Ep. 18 | 52% | 53.0% |
| Ep. 19 | 82% | 68.3% |
| Ep. 20 | 42% | 58.7% |
| Ep. 21 | 69% | 64.3% |
| Ep. 22 | 83% | 64.7% |
| Ep. 23 | 97% | 83.0% |
| Ep. 24 | 82% | 87.3% |
| Ep. 25 | 87% | 88.7% |

- **Music/silence agreement: 0.82** with the composer's final mixes.
- **Scene boundaries: F1 0.80, precision 0.91, over 21 episodes.** In plain terms: it finds about 7 in 10 scene changes, and when it marks one it is right 91% of the time.
- **Speed:** a full run on a 35-minute episode takes 34.7 minutes (previously 50), i.e. faster than the episode's running time. Music regions are written to the session at minute 21.3 of the run (previously 49.5). 35-minute broadcast episode, Oct 2026, in-house runs on the studio's GPU worker.
- **Same melody repeated within 10 minutes, per episode** (same drama, Sept–Oct 2026):

| Episode | Repeats |
|---|---|
| Ep. 13 | 6 |
| Ep. 14 | 10 |
| Ep. 15 | 7 |
| Ep. 16 | 4 |
| Ep. 17 | 2 |
| Ep. 18 | 1 |
| Ep. 19 | 2 |
| Ep. 20 | 6 |
| Ep. 21 | 3 |
| Ep. 22 | 2 |
| Ep. 23 | 0 |
| Ep. 24 | 1 |
| Ep. 25 | 2 |

### Music ID

In-house bench, 3 Oct 2026: 7 broadcast episodes, about 3.8 hours of music, real dialogue mixed in at each level.

| Condition | Identified correctly |
|---|---|
| Music without dialogue | 99.1% (up to 99.9% on later episodes) |
| Dialogue on top, same level (0 dB) | 97.9% |
| Music 10 dB under dialogue | 95.2% |
| Music 20 dB under dialogue | 86.3% |

- Without source separation, results are within 0.6 points of the above everywhere except music 20 dB under dialogue, where accuracy is 78.2% (vs. 86.3% with separation).
- Real editor reference cuts with dialogue: 96.9–98.8%.
- Entry timing: 98% of entries placed within ±0.5 s (clean music).
- Speed: about 3 minutes per episode.

## Honest limits

- Music ID does not name music that is outside your library.
- Very quiet music (about 17 dB or more under dialogue) can be missed.
- Exits under dialogue are placed less precisely than entries.
- The measurements are in-house, on the studio's own deliveries; the ScoreFlow keep-rate figures come from one 90-episode broadcast drama.

## Who it's for

- **Streamers** — music placed and cleared across whole seasons, with cue sheets ready for every territory.
- **Broadcasters** — every episode spotted and edited on deadline; cue sheets straight from the broadcast copy.
- **Post-production houses** — the music editor's first pass in minutes, open anywhere from a private link.
- **Production companies** — music decisions made with the picture, not after it.
- **Ad agencies** — the right library track cut to the spot, with rights you can trace.
- **Music libraries and rights societies** — know where every track played, even under dialogue.

Starting where music is hardest: long-running TV drama.

## Roadmap

- **Original music without generative audio:** Claude writes the MIDI, licensed instrument libraries play every note in the DAW, and the system bounces its own mix. Because no audio is generated, the rights stay clean for broadcasters.
- **Local-first DAW with cloud sync on demand**, opened from a private web link (see "The DAW" above).

## Track record: Batuhan Fırat, Founder & Producer

Fifteen years of making music for TV and brands. Now rebuilding the whole pipeline around AI.

- 66 film & TV productions, 39 brand projects, making music since 2011.
- Music producer for TV; fifteen years running JingleTV, the studio's music brand for custom music (https://www.jingletv.com/). Co-founder of Soundoo, Inc. (Delaware).
- Series & film credits include: Sustalı Ceylan, Yalandan Sevgilim, Alparslan, Bozkır.
- Brands include: Coca-Cola, Turkish Airlines, Facebook, Ford, beIN SPORTS, Fanta, League of Legends: Wild Rift, Royal Match.
- Music for award-winning work (the awards belong to the projects):
  - Hagia Sophia History & Experience Museum — two MUSE Gold Awards (2024) and a MONDO-DR Award (2025)
  - Selsil Ninja — 3 Kristal Elma and a MediaCat Felis
  - Tadım Bar — Effie Türkiye 2018 Gold & Silver
- Full portfolio: https://batufirat.io

## History

11,720 commits since February 2026 — every single week. 4,465 on the core pipeline (video analysis, music analysis, scoring, DAW) · 6,385 on the systems that feed and speed it up · 870 on creative AI (storyboard & AI video). Unrelated projects excluded.

Commit counts cover all of Batuhan Fırat's repositories since 1 February 2026, deduplicated by commit (ISO weeks, Monday start; the latest week is partial). The core pipeline is video analysis, music analysis, scoring/ScoreFlow and the DAW. Supporting systems are the tools that feed and speed it up, such as agent orchestration, computer-use automation, memory, the app shell, music production and communications. Creative AI covers the storyboard and AI video tools. Unrelated projects are excluded.

- 7 Feb 2026 · Video analysis: First commit of the video analysis engine: where the project began.
- 10 Feb 2026 · ScoreFlow: First commit of ScoreFlow.
- 15 Feb 2026 · Music analysis: First commit of the music analysis engine.
- 19 Feb 2026 · ScoreFlow: Unified pipeline spec connecting ScoreFlow with video and audio analysis.
- 21 Feb 2026 · ScoreFlow: First end-to-end run of the full pipeline: 161 library tracks, 21 cues placed.
- 22 Feb 2026 · ScoreFlow: First automatic cue-sheet bundle.
- June 2026 · All engines: Rebuilt as native engines: picture, scoring, music understanding and the DAW.
- 21 Aug 2026 · ScoreFlow: New ScoreFlow DAW codebase.

## Recent changes (lab log)

- 7 Oct 2026 · ScoreFlow: The DAW keeps detailed waveforms while you navigate the timeline.
- 6 Oct 2026 · ScoreFlow: The Spotter starts from per-scene drafts while the scene book is still being verified: music regions land about 28 minutes sooner.
- 5 Oct 2026 · ScoreFlow: A Linux GPU worker joins; face analysis runs locally on the GPU.
- 4 Oct 2026 · ScoreFlow: Learning loop: after every delivery, the Scorer compares its picks with the composer's final mix and learns from the composer's edits.
- 4 Oct 2026 · Music ID: Searches the whole library when the source is unknown; one runbook is now the single source.
- 3 Oct 2026 · ScoreFlow: The Scorer plans the whole episode's tracks at once: repeats inside 10 minutes dropped from 8 to 0, and edit points land closer to the cut (0.64 s on average, from 1.02 s).
- 3 Oct 2026 · Music ID: Reads full episode mixes: which track, where, from which source second and how it was edited. Benchmarked on our own deliveries with real dialogue mixed in at 0, −10 and −20 dB: 99.1% without dialogue.
- 28 Sep 2026 · Music ID: 36 commits in one day on identification quality.
- 27 Sep 2026 · Music ID: A reproducible quality foundation: every change is measured against the same bench.
- 26 Sep 2026 · Music ID: First private pilot: verifying edited tracks and their source-time intervals.

## Contact

- Email: batuhan@pelandbee.com
- Investors: deck and a 15-minute call.
- Broadcasters and studios: send one episode, get the spotting draft and cue sheet back.
- LinkedIn: https://www.linkedin.com/company/pelandbee/
- Instagram (media channel): https://www.instagram.com/pelandbee/
- Custom music: JingleTV, https://www.jingletv.com/