V26.06: i trained a 125m model to autocomplete piano on-device

Overview

This project delivers a novel and genuinely useful creative tool. The "Copilot for piano" analogy is strong and immediately understandable. The on-device execution is a significant competitive advantage, offering real-time feedback, privacy, and offline functionality—solving the latency and cost problems that plague cloud-based AI music tools. It lowers the barrier for musical inspiration and could be a valuable practice or composition aid for a wide range of musicians, from hobbyists to professionals. The technical execution is impressive.

Score Breakdown

What They Do

I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.

Why It Matters

Business Model

To be determined based on project analysis.

CRP Decision Report

🧠 Advisor

[score: 8]

[score: 8.0]

🧨 Devil

The risks are significant and must be carefully weighed:

Model Quality Ceiling: A 125M parameter transformer is compact. While fast, it may lack the long-term musical coherence, harmonic understanding, and stylistic diversity needed to be a serious composition tool rather than a novelty. The output might feel repetitive or shallow over time.
Market Viability: The intersection of "AI enthusiast," "pianist with a MIDI keyboard," and "willingness to pay for a subscription" is a very small niche. Converting a free, impressive demo into a sustainable business is a notoriously difficult challenge.
Platform & Competitive Risk: On-device music AI is a natural feature for Apple to integrate directly into GarageBand or Logic Pro, leveraging their own Silicon and Core ML. This would instantly make a standalone app obsolete. The open-source community is also rapidly converging on this problem.
User Retention: The initial "wow" factor of AI autocomplete can wear off quickly. If the model doesn't continuously surprise and delight, users will abandon the app, especially if the musical suggestions fail to meet their creative standards.

[score: -7]

[score: -7.0]

📚 Historian

Historical patterns for AI music generation for consumers are largely cautionary. Startups like Jukedeck, Amper Music, and Melodrive struggled to find sustainable product-market fit in the consumer space, often pivoting to B2B or being acquired and shut down. The "Copilot for X" pattern is historically very strong in coding (GitHub Copilot, Tabnine), but music is inherently more subjective and less deterministic than code. The on-device, real-time execution is a favorable divergence from the cloud-based latency that killed past consumer experiences. This specific execution pattern is unique enough to merit cautious optimism, but the historical business model data for consumer AI music is deeply negative.

[score: 2]

[score: 2.0]

🧮 Budget Steward

The project is already at the MVP stage with a live free app. The cost to get to the first paying customer is exceptionally low. The remaining work is primarily business logic and monetization:

Developer Account: $99/year (App Store)
Payment Processing / Subscription SDK: $0 (RevenueCat or similar offers generous free tiers)
Landing Page / Support Site: ~$10/month for 3 months
Legal Templates: ~$100 (Terms of Service, Privacy Policy) Total estimated cost to reach the first paying customer is approximately $250. This is well under the $1,600 threshold.

[score: +1]

🧱 Founder

This is a brilliant technical achievement and a fantastic portfolio piece. The project is already built, which removes the highest risk of an indie project. The "Should we build this?" is a definitive yes—it's already done. The "How?" involves adding a simple monetization layer (e.g., a subscription for advanced features, longer generations, or a specific model pack) and a focused marketing push to the piano/music production community. The timeline for implementing a subscription paywall and shipping the monetization update is 1 month. The risk is the market size, but the low cost of entry and the strength of the demo make it a worthwhile gamble. It is a strong pass.

[score: 7]

[score: 7.0]

Why This Made Only 10

This project was selected because it scored highly across our evaluation framework. The CRP analysis confirmed strong potential, making it one of the most promising opportunities this week.