MixLab: AI-Powered Mix Curation for DJs

The Problem

There is a specific kind of paralysis that comes with having too much music.

A DJ who has been buying records consistently for years will end up with a collection that eventually exceeds what they can hold in their head. At 500 tracks it is manageable. At 1,000 you start to forget things. At 2,500 you are effectively flying blind. You know the records you reach for out of habit. Everything else sits in the crate, purchased and tagged and never played.

The instinct is to keep buying. The real problem is curation.

Creativity has the same problem. Building a coherent set is a genuine act of musical reasoning: sequencing tracks, managing energy from opener to closer, deciding when to surprise and when to sustain. It also corrodes with familiarity: a DJ who knows their records well tends to reach for the same combinations, the same trusted paths through the collection. MixLab works on both. It identifies what has not been played, and applies a structured creative process to that unexplored material.

Where It Sits in the Ecosystem

MixLab is the second tool in a two-part workflow that begins with TuneFinder.

TuneFinder answers the question: what should I buy? It monitors six release platforms weekly, scores candidates against a behavioural taste profile derived from mix history, and surfaces 15-20 relevant tracks.

MixLab answers the question that comes next: what should I play, from what I already own?

TuneFinder populates the Rekordbox library. MixLab curates it. Together they form a closed loop: discover, collect, curate, perform.

System Architecture

The pipeline runs through several distinct phases on each execution.

Phase 1: Collection Parsing

Rekordbox exports its library as XML. MixLab reads this file directly using lxml, pulling out artist, title, BPM, key (in Camelot notation), genre, and track ID. Any track missing BPM or key data is excluded, since harmonic sequencing requires both. One known edge case is handled automatically: some Drum & Bass tracks get tagged at half their true BPM in Rekordbox. The system detects these by genre and doubles the value.

Phase 2: Play History Cross-Reference

The Changsta API holds a paginated catalog of every track played in a recorded mix. MixLab pulls this down in full - typically 1,200+ entries spanning a decade of mixes - and filters the collection against it. Matching uses normalised comparison: unicode normalisation, featured artist stripping, punctuation removal. This is necessary because track titles often differ slightly between the Rekordbox tag and the mix catalog entry (feat. vs ft. vs featuring, that kind of thing). The result is a filtered pool of genuinely unplayed tracks, grouped by genre.

Phase 3: Collection Overview and Genre Clustering

Before any genre-specific work begins, MixLab prints a full availability table across the entire collection. For each genre, it shows how many tracks are in the library and how many are unplayed, rendered as a simple bar chart in the terminal and Discord report. This gives an at-a-glance picture of where the collection is concentrated and which genres might need more attention or buying. A genre sitting at 95% played is telling you something different to one sitting at 55%.

The report also surfaces a separate section of outliers - tracks whose Rekordbox genre tag does not map to any known label. These are not discarded quietly. Seeing them listed is useful in practice, because it often reveals tagging inconsistencies that have crept in over years of importing from different sources. A track tagged "House;Downtempo;Leftfield" or left with a blank genre is easy to miss in Rekordbox itself; surfacing it here makes it easy to go back and fix. Getting the tags right means those tracks will be picked up correctly on the next run.

The genre map itself translates Rekordbox's inconsistent free-text tags into consistent labels, so house covers eight distinct variations. Within each cluster, a BPM median filter removes any tracks sitting more than 6 BPM outside the group median. Outliers that share the requested genre character (four or more) are passed to the LLM pipeline as a "Misc" pool rather than dropped entirely.

Phase 4: Harmonic Pre-Sort

Before anything hits an LLM, tracks in each cluster are pre-sorted by Camelot wheel compatibility. The algorithm walks the list: from the first track it finds the next harmonically compatible one (adjacent or same position on the wheel, same or relative mode), preferring the lower BPM candidate where there is a choice. By the time the LLM receives its input, it is already a musically coherent sequence rather than an arbitrary list.

The Two-Stage LLM Pipeline

The two stages do different jobs, and that distinction matters for cost as much as for quality.

Stage 1: Technical Shortlisting (Provider Cascade)

The first stage is purely analytical. Given up to 40 tracks, the model groups them into 2-3 candidate pools of 15-25 tracks each, based on a shared BPM centre and harmonic character. It is told explicitly not to make ordering or editorial decisions - its job is pre-screening, not curation. The output is a JSON array of shortlists with rough descriptive titles ("Deep 122 BPM / 4A-7A Pool") and a one-line sonic mood.

This stage runs against a provider cascade: MiniMax M2.7 first, then Groq (Llama 3.3 70B), Gemini 2.5 Flash, OpenRouter (DeepSeek), and Claude Sonnet as a last resort. The cascade prioritises cost and speed. MiniMax and Groq are fast and cheap; the Anthropic fallback exists to keep the pipeline running if everything else fails, not because it is the preferred choice here. If a provider times out or returns malformed JSON, the next one is tried without any intervention needed. MiniMax's reasoning models emit inline <thinking> blocks that get stripped before JSON parsing.

Genre pools larger than 40 tracks are chunked and processed in batches, with the resulting shortlists merged at the end.

Stage 2: Creative Curation and Narration (Anthropic, fixed)

The second stage is where the actual curation happens. All candidate shortlists from Stage 1 are sent - capped at six, ranked by pool size - to Claude Sonnet in a single call. This stage has no fallback. The quality of output here justifies both the cost and the hard dependency.

The system prompt instructs the model to work as a mix curator. For each shortlist it has to select the best 8-12 tracks from the pool, sequence them as an intended play order, design a deliberate energy curve across the set (tension, build, peak, release, landing), and justify the editorial choices it makes. The instruction is explicitly not to optimise only for BPM and key compatibility. The goal is flow, memorability, and emotional payoff. Bigger key jumps and tempo pivots are encouraged when they serve the journey.

Each concept in the output includes:

• A creative name, not the pool label from Stage 1
• Track order with Camelot key and BPM annotated inline
• An arc description covering the emotional journey, structural logic, and specific moments worth naming
• Opener rationale: why this track works first
• Closer rationale: why this track works last
• Standout transitions: 1-3 specific moves called out with precise context
• Assumptions: what was inferred from metadata where audio analysis was not possible

The prompt specifies the tone as opinionated, musical, honest, no marketing language, no filler. The outputs reflect that. A concept called "Slow Burn" opens with Alan Dixon's La Danza and explains in plain terms exactly why the key jump from 4A to 7A at the midpoint earns its emotional shift. "Pressure Front" names the Trevino remix of Goldie's Inner City Life as "a house music weapon" and accounts for why it holds up at 127 BPM within the 8A cluster. The reasoning is specific to the actual tracks in the pool, not generic advice dressed up in music-sounding language.

Prompt Engineering

The hard part of writing this prompt had nothing to do with LLM mechanics. It was deconstructing what DJs do. What does an opener do that a second track cannot? What makes a closer feel like an ending rather than just the last thing that played? Why does a harmonic jump that looks wrong on paper sometimes feel inevitable on a dancefloor? These are not questions DJs typically answer. They just play.

Getting useful output meant turning that instinct into explicit rules, without killing the creativity in the process. Most AI music prompts fail here: they describe genre characteristics and energy arcs in the abstract, and get back something technically correct and musically inert. The prompt evolved through that specific failure, each iteration stripping one more layer of assumption to find the logic underneath.

A ceiling remains. Stage 2’s output quality is bounded by the LLM’s world knowledge of individual tracks. The Assumptions section makes this visible; it does not fix it. The obvious next step is richer input: cue point data as a proxy for intro and outro length (it is already in the Rekordbox export), energy ratings, colour tags. Less inference from a track name, more signal from the file.

Output and Delivery

Rekordbox XML Export

The curated track IDs from Stage 2 are matched back against the raw Rekordbox XML. MixLab generates a single merged export file with a shared track collection and all concepts as named playlists, nested inside a dated folder ("Mix Lab - house - 2026-03-15"). Load it into rekordbox and each concept appears as a ready-to-use playlist in the intended play order.

Discord Delivery

The full text report and the Rekordbox XML are sent to Discord via webhook. The report arrives as a formatted message and the XML as a file attachment. The output is on any device immediately, archived in the channel automatically, with no need to touch a terminal to retrieve it.

Results

A house genre run, in numbers:

• 2,574 tracks parsed from the Rekordbox library
• 1,224 played tracks cross-referenced from the Changsta API
• 336 unplayed house tracks identified across 8 Rekordbox genre tags
• 6 named, sequenced mix concepts generated, each with 10-12 tracks
• Total runtime: 8 minutes 9 seconds, most of which is Stage 2 LLM time

Six concepts, genuinely distinct in character: from a jazz-inflected deep house arc built around patience and warmth to a high-pressure peak-time set anchored by Inner City Life and The Blessed Madonna. Each draws entirely from music that is owned but unplayed. Each loads directly into professional DJ software.

The creative brief, the arc, the transition rationale: these are starting points, not instructions. Which concepts are worth pursuing, which tracks to swap, how a set unfolds on the night: those stay human decisions. MixLab expands reach into the collection. The taste required to use it is still yours.

A useful edge case: when Stage 2 identifies a concept where the unplayed pool is too thin for a proper set, it says so directly and flags how many tracks are missing. That closes the loop. MixLab points at what to go and find.

In the Booth

I took the Slow Burn concept and played it as an actual mix. These are my honest thoughts.

What Worked

The overall journey held up. The set had a real shape to it and I enjoyed playing it. What I got most from it was tracks I had completely forgotten about, which is the whole point of the tool.

Where It Missed

Reading tracks from their names. Alan Dixon’s La Danza and DJ Rocca’s Como Esta were described as Latin-tinged and percussion-forward. In practice, La Danza is quite psychedelic and Como Esta is more warehousey. The AI almost certainly picked that up from the track names, which both sound Latin. It had no way to know what they actually sound like. This is exactly the limitation the Assumptions section exists to surface, and it is a direct argument for enriching the Track model with better metadata. A genre tag or energy rating would have caught this. A track name cannot.

The vocal clash. The Felix Da Housecat Röyksopp remix into the Kraak and Smaak Yuksek remix had clashing vocals, something you would never know from a tag or a title. The prompt now explicitly flags this risk in the Assumptions section when arrangement data is unavailable, but flagging a risk is not the same as resolving it. In practice it forced an interesting solution: using STEMs to isolate and blend the two vocal layers, which ended up working quite nicely. A limitation that led somewhere good.

Getting the closer wrong. Saison’s We Are the Machines was predicted to be a driving, machine-like closing statement. It is a chilled, vibey tune. It worked as a closer, but for reasons the AI did not anticipate. This is the world knowledge ceiling in action: the prompt can define what a closer needs to do. It cannot verify whether a specific track does it.

Something I Did Not Expect

I skipped Marea by Fred Again and The Blessed Madonna. Not because it was in the wrong place (the sequencing was fine) but because listening to it in context I realised I was not into it anymore. MixLab surfaced it because I had never played it in a recorded mix. Useful, just not in the way I expected: it flagged a track that probably should not be in my collection anymore. The tool found a gap I did not know I had.

Tech Stack