Inspiration
My larger project is called Medusa. Medusa is an AI music intelligence platform I've been building to give independent artists a complete picture of their own music. SyncSnake is one of Medusa’s many snakes. It goes out into the world and hunts for a variety of opportunities that exist for musicians but may feel out of reach for many.
The inspiration came from watching artists I know pour everything into their work, upload it to streaming, and wait. Not because streaming is wrong. Streaming platforms are where we all listen, and the path there is clearly marked. But sync licensing, grants, festival placements, competitions, music libraries, and ad agencies are viable paths too. They're just harder to find. For an independent artist, tracking them down can feel impossible. SyncSnake exists to overcome at least part of this hurdle.
What it does
SyncSnake is a multi-agent system that hunts the open web for sync licensing opportunities, grants, festival calls, music supervision contacts, and even advertising agencies. It puts these opportunities within reach of the artist.
It takes a track profile from Medusa (genre, placement profile, mood, the full fingerprint) and deploys five specialist agents in parallel, each expert in a different corner of the industry:
- Sync agencies
- Music supervisors
- Grant platforms
- Licensing briefs
- Direct-contact research
They run simultaneously on Google's Agent Development Kit (ADK), grounded in live Google Search.
Results come back as a swipe deck — a familiar card-based UI where you can flip through a large set of leads quickly, saving what's relevant and passing what isn't. Less like a spreadsheet, more like something you'd actually want to use after a long recording session.
And because these agents are researching the live web, there's a real problem: how do you know they're not making things up? That's where Arize Phoenix lives in the stack — every run is fully traced (spans, latency, token cost per agent), and every result is scored for faithfulness against its source. Anything that doesn't hold up gets filtered before it reaches the user.
How I built it
The agent layer is built on Google ADK with Gemini as the underlying model, grounded in Google Search for live web access. Five specialist sub-agents fan out from a Scout root span, each returning structured leads that get merged, deduped, and scored.
Observability is wired in via Arize Phoenix using OpenInference spans. Every run exports a full trace:
- Root Scout span
- Specialist branches
- Latency and token cost per call
A faithfulness evaluation pipeline scores each returned lead against its cited source; low scores (we saw real 0.19s in production) get flagged and filtered. Those failures accumulate in a Phoenix dataset automatically, building a regression set for future fine-tuning. Anti Gravity has been indispensable throughout this process.
Challenges I ran into
The honest answer: trusting agents can be difficult. Early runs came back with results that looked great on the surface, but from other challenges with Medusa, I knew that things are not always as they first appear. When you're handing leads to an artist and launching a SaaS for artists, accuracy is incredibly important.
Getting Arize's faithfulness evaluations wired into the pipeline — and actually acting on the scores, not just logging them — was the thing that made the system feel trustworthy instead of just impressive.
The other challenge was scope. Medusa is a big system, and SyncSnake is one piece of it. Keeping the demo honest about what SyncSnake actually is (one snake, not the whole head) while still showing enough of Medusa to give it context — that's a design and storytelling challenge.
Accomplishments that I’m proud of
I'm not a trained engineer — I came to this from the design side of music. So when I say I learned the difference between ADK and GenAI API calls, or that I finally understood how a trace works and why it matters, those aren't small things.
But the moment that hit hardest was midway through the project. I looked at my git repo and said to a friend:
"Wow, I can actually understand what these files are in git now."
He responded:
"Wow, that must be like seeing through the Matrix for the first time."
He wasn't wrong.
Beyond the personal: I'm proud that the observability layer isn't decorative. The faithfulness scores changed the output. Real hallucinated leads got caught. The dataset grew. That feedback loop working in production — not just in a demo — felt like the system actually earning the trust we will be asking artists to place in it.
What I learned
Agent reliability and agent capability are two different problems, and they require different tools. You can have a highly capable agent that you still can't hand to a real user because you have no visibility into what it actually did. Arize closed that gap — not just as a monitoring layer, but as a way to build trust in the output over time. Every flagged example is a data point toward a system that gets better.
Also: five parallel agents is a great idea until one of them goes rogue on token cost. Tracing per-call spend was a revelation.
What's next for SyncSnake
On the product side, it will be interesting to test out SyncSnake moving forward — same architecture, new domains. Eventually, SyncSnake talks back to Medusa: the opportunities she finds should inform what the platform recommends the artist do next.
Right now, I'm moving toward beta. Honestly, once real artists are using it, SyncSnake's roadmap becomes theirs. I have a vision for what she should do — but the people actually trying to get their music placed will know better than I do what she's missing. I'm building toward that conversation.
The snakes are just getting started.
Log in or sign up for Devpost to join the conversation.