Wavelength

Inspiration

Navigating conversations is usually intuitive. For millions of neurodivergent people, the unwritten rules of networking are invisible.

Wavelength came from watching someone we care about hit that wall. Someone close to us is autistic and as sharp as anyone we know. Not long ago we watched them introduce themselves to a professional at a networking event, and early on they mentioned they were autistic. It was honest, but the timing caught the other person off guard. He tried to roll with it, the conversation went stiff, and it fizzled out within a minute. They felt the tension but could not read where it came from. It knocked their confidence badly.

That is the gap Wavelength fills. They knew something had gone wrong, but nobody ever tells you what it was. We built the missing translator.

What it does

Wavelength is a private social coach for neurodivergent people. During a live conversation it nudges you when you miss a cue, and afterward it breaks the whole interaction down so you can learn from it on your own time.

The loop is simple. You give explicit visual consent first, and recording only ever starts when you start it. While you talk, a quiet dashboard follows the conversation. If engagement drops, you get a gentle nudge with the reason attached, like a suggestion to ask a question because you have been talking for over a minute. When you are done, you get an AI debrief: what went well, where it slipped, and specific things to try next time.

How we built it

Wavelength is a React frontend and a Node Express backend, deployed end to end on DigitalOcean.

The frontend uses MediaPipe to read facial cues and body language, and it runs entirely on the device. The raw video and audio never leave the laptop.

The Node Express backend runs on DigitalOcean's App Platform, where it serves the React client and acts as a secure proxy. DigitalOcean Gradient AI Inference powers both the live reasoning and the debriefs, and because it is one API behind one key, we could switch between Claude Haiku and Claude Sonnet without touching the rest of the code. DigitalOcean Managed Postgres stores the conversation transcripts, so the app can pick up on recurring patterns over time.

Challenges we ran into

The hardest problem was an ethical one: how do you get the benefit of AI perception without building something that spies on people. Our answer was a hard boundary. The raw video never leaves the browser. We turn it into derived numbers many times a second and send only those numbers and the text transcript to the backend.

The interface took real care too. Our user is already spending energy just to follow the conversation, and a second voice whispering advice in the moment splits their attention and makes things worse. So we tuned the live nudges to be rare, plain, and easy to ignore, and we built the AI to suggest rather than diagnose. It never claims to know what someone feels. It only points to what it can actually observe.

We also changed direction under the hood more than once. We started on a Python backend and moved to a Node monorepo. Before writing the core logic, we wrote a test script to verify our model access keys, work around an SSL issue on the Postgres connection, and confirm our tool-calling was solid.

Accomplishments that we're proud of

We got a live perception engine working: it reads webcam data in real time, keeps every raw frame on the user's device, and still routes enough signal through DigitalOcean to produce coaching that is both useful and kind.

What we learned

The best AI feature here turned out to be restraint. We assumed real-time coaching would be the heart of the product. Testing with our actual audience showed the opposite: whispering in someone's ear during a conversation makes the anxiety worse, not better. The value is in the calm retrospective afterward, when there is space to actually take it in. We also learned how much simpler it is to run an entire backend stack on a single platform.

What's next for Wavelength

Next is memory. We want to deepen the long-term store in Postgres so Wavelength can follow someone's progress over months and surface the habits that only show up across many conversations. There is also a clear market with social skills therapists, who could give clients a way to practice between sessions.