Music Company Internal Tool
A music business asks the same question every morning: what deserves attention today. This tool answers it from a record it keeps itself, and a studio then turns that same answer into a finished vertical clip on the company's own hardware.
- DomainMusic business, artist and repertoire
- EngagementInternal tool build, data platform and video studio
- StageRunning in daily use
A record it keeps itself
Chart positions and, where a source publishes them, absolute listen counts are collected on a schedule from many major platforms and storefronts, and every one of those daily readings is kept. Each row is stored with the full original payload behind it and a tag recording how it was obtained, so any answer the tool gives can be traced back to the exact response that produced it. A day lost to a restart is not a hole in the record, because the run catches itself up, and a source that fails is written to an error table while the rest of the snapshot completes.
The second half is a studio that turns one of those answers into a finished clip. It runs as a staged pipeline: find the track's most replayed window, cut there, separate the vocal, time the words, correct those words against the real lyrics, generate the scene imagery, and render vertical with animated text. It runs on the company's own GPU hardware, and that constraint shapes the render decisions. This is an internal tool for a single operator rather than a product, and the architecture is honest about that.
A picture assembled by hand, every morning
Everything needed to decide what to work on was published somewhere. None of it was in one place, and none of it was kept.
The morning started by opening one dashboard after another
Positions were read off one dashboard after the next, one storefront at a time, and copied into a sheet. The picture was complete around the time it stopped being current, and the next day the same work started again from nothing.
A position on its own says nothing
Without a kept record of previous days, a track climbing fast and a track sitting still looked identical on the page. Movement existed only in whoever happened to remember last week, and it could not be asked about after the fact.
Acting on a decision cost a day of editing
Once a track was worth doing something with, the publishable version meant finding the right seconds by ear, transcribing the words, timing them by hand and assembling the video in an editor. That cost is the reason most decisions never became anything.
Both halves of the desk
A data side that keeps the history and ranks movement over it, and a studio side that turns a decision into a finished clip.
Scheduled snapshots into a time series
The tool takes its own reading every day, on a schedule, across many major platforms and storefronts, and keeps all of them, so the history grows on its own.
- Chart positions across many major platforms and storefronts, with absolute listen counts wherever a source publishes them.
- Every row kept with its full original payload and a quality tag, so an answer can be traced back to the response that produced it.
- A failing source is recorded and skipped, the run carries on, and a missed day is caught up automatically.
Momentum and breakout detection
The history is what makes the tool worth having. Movement is computed over it in the database itself, at several distances back, then ranked and scored per genre and storefront.
- Rank movement measured at several distances back, so a climb is visible before it becomes a position.
- A momentum board that ranks on acceleration rather than totals, on whether a gain is speeding up or cooling, with several weighting profiles and time windows.
- A breakout detector that flags a candidate only when independent conditions agree at once, and drops it as soon as one of them is missing.
The studio as a staged pipeline
A track goes in and a finished vertical clip comes out, as a sequence of stages that each hand their result to the next and record what they decided along the way.
- The clip is cut at the track's most replayed window, and the opening is discounted because every replay restarts there.
- The vocal is separated before transcription and the words are timed, then the words themselves are corrected against the real lyrics so a mis-heard line cannot reach the render.
- Scene imagery is generated for each lyric window and animated in short segments, then cut together and rendered vertical with animated text, with and without the music.
One dashboard in front of all of it
A single surface for both halves, built so the operator can see what the data says and what the machine is doing at the same time.
- Chart drill-downs, a rank ladder per track and a dense heatmap view, all driven from the address bar so any view can be handed over as a link.
- The pipeline is started from the dashboard with its progress streaming stage by stage, and a run history underneath it.
- A timeline editor for word by word timing, drag to position overlays against the vertical frame, and a live gauge of what the graphics hardware is holding.
Decisions still visible in how it behaves
Naming the decision is more useful than naming the library, and each of these is still readable in the way the tool answers.
A decision record the plans defer to
Every architectural decision that shaped the system is written down with the consequences it forces, and when one was later reversed the original entry kept its place with the reversal and its reason appended underneath. If a plan and the record disagree, the record wins. That rule is the only thing that stops a written decision from quietly becoming fiction a month later.
One join key, with fuzzy matching banned in writing
Matching the same recording across sources uses the industry standard recording identifier and nothing else. Falling back to matching on title and artist text is ruled out in the decision record, because that kind of match invents agreement that was never in the data. Where the identifier is missing, entries stay apart, so the tool understates agreement between sources.
A slow query fixed, with the reasoning left in place
The detection query read the same heavy join several times over and had grown slow enough to be felt at the desk. Computing that join once and reusing it turned a wait into an answer. The measurement, the issue it came from and the reason the query is shaped that way sit in a comment beside the change, so the next person to open the file does not have to rediscover any of it.
Windows that degrade instead of refusing
Ask for a long window while the archive is still short and the tool answers over the history it actually has, says so, and grows into a true window as the record fills. In the same spirit, the weighting behind the momentum score is annotated as reasoned judgement rather than something fitted to labelled examples, with the backtest deferred until there is enough history to run one honestly.
What changed at the desk
Qualitative, and deliberately so. Nothing here was measured against a baseline, so nothing here is claimed as a number.
Kept
A record instead of a snapshot
The picture stops resetting every morning. Each day's positions and counts stay, so a question about movement can be asked after the fact instead of depending on who remembers last week.
Earlier
Attention lands while a track is still climbing
Ranking on whether a gain is speeding up moves attention forward, towards tracks still climbing rather than ones already at the top of a list everyone else is reading too.
Repeatable
Publishing stops being the expensive part
The steps that used to be an afternoon of manual work run as one pipeline on hardware the company already owns, so the cost of acting on what the data says no longer decides whether it gets acted on.
Is your morning question spread across too many dashboards?
If the answer you need already exists in several places and somebody assembles it by hand every day, that is usually one internal tool. Tell us what the question is and we will tell you what it takes to answer it on a schedule.
Discuss your project