A&R Radar

A&R Radar scans new SoundCloud uploads, uses HumanStandard to filter out AI-generated music, and ranks the human artists by real listener traction so scouts hear music worth signing.

Comment

Inspiration

A&R scouts (the people at labels who find artists to sign) spend hours combing SoundCloud and similar platforms for new talent. Two things now waste that time: AI-generated tracks flooding upload feeds, and play counts that can't be trusted at face value. An "artist" that doesn't exist, or a track with inflated numbers, isn't a lead. We built A&R Radar for the HumanStandard track so a scout only sees music worth hearing.

What it does

A&R Radar is a four-stage pipeline with a web dashboard:

  1. Scout. Finds fresh SoundCloud uploads from emerging artists. It skips DRM-locked tracks, DJ mixes, covers, remixes, accounts with large followings, and channels that post hundreds of tracks.
  2. Forensic. Sends each track's audio to the HumanStandard API. AI-generated tracks are discarded and logged. Borderline verdicts are flagged for a person to review, so a real artist isn't binned on a shaky call.
  3. Analyst. Measures real traction: likes, reposts and comments per play, audience size, and play-farming patterns.
  4. Outreach. Produces an A&R score and a scorecard for every lead, with HumanStandard's evidence explained in plain language, plus optional Slack or Discord alerts.

The dashboard shows every scanned track with its verdict, score, evidence and a "how this score is calculated" breakdown. You can filter, sort and shortlist leads.

How we built it

  • Python for the pipeline: yt-dlp and ffmpeg to fetch audio, httpx for the APIs, and SQLite for results.
  • SoundCloud's search API for discovery, which gives upload dates and follower counts.
  • The HumanStandard API (asynchronous jobs that we poll) for the human-versus-AI verdict, its confidence, a risk timeline across the track, and nearest reference recordings. We built and tested against its free mock mode first, then ran real scans.
  • A local dashboard built with Python's standard-library web server and plain HTML, CSS and JavaScript, with no framework.
  • Built with help from Claude Code.

Challenges we ran into

  • Search isn't ordered by recency. Our first live run returned nothing, because SoundCloud search ranked years-old tracks first. We switched to a date-filtered query that also returns follower counts.
  • Limited credits. Each scan costs one HumanStandard credit, so we rank candidates by traction first and scan only the best, one track per artist.
  • Noisy results. Searches pulled in covers, remixes, fan re-uploads and radio-style channels. We added filters for those, though title matching is crude and will sometimes skip a real original.
  • Reading detector output. A "human" verdict at 0.49 confidence looks contradictory until you learn that 0.5 is a coin flip. We rewrote the evidence into plain language.
  • Small numbers. Engagement rates on tracks with a handful of plays are noise, so we damp them.

Accomplishments that we're proud of

  • A working end-to-end pipeline that ran on real, current SoundCloud uploads, about 20 real scans in all.
  • Results a scout can understand: every score shows exactly where it came from, and every detector field is explained.
  • A cautious design. AI tracks are auto-discarded only when the strictest signal agrees, and borderline cases go to a human.
  • A dashboard and scorecards we're happy to put in front of a scout.

What we learned

  • Play counts say what's popular, and a detector says what's real. You need both to find music worth signing.
  • A detector's output is a probability, and the honest way to use it is to keep a person in the loop for borderline calls.
  • Data quality drives everything. Most of our work went into filtering and ranking, not the AI call itself.
  • Our A&R score is a heuristic we designed, not a validated predictor, and we say so in the app.

What's next for A&R Radar

  • More sources. TikTok, Spotify and Bandcamp, since real scouting isn't only SoundCloud.
  • Growth over time. Track how fast an artist is rising week over week, not a single snapshot.
  • Validate the score with working A&R professionals and tune it to the signals they actually trust.
  • An investigator agent that looks across an artist's uploads for patterns, such as several AI-flagged tracks from one account.
  • Stem-level checks using HumanStandard's hybrid analysis for borderline tracks.
  • A hosted version with scheduled scans, so scouts get a weekly digest of new human-made leads.

Built With

Share this project:

Updates

Submission history