Inspiration
I've followed CDawg's Cyclethon every year since the first one. Seeing Ironmouse pop in through an iPad on a Discord call got me thinking about another way friends could join an adventure remotely.
That idea became Slipstream: a way for someone at home to ride alongside the person who's out in the world, with a view they can look around and an avatar they can make their own. They could pull up beside a friend, fall back, coast while chatting, or mess around together along the way. The aim is for joining in to feel immersive, intuitive, and fun.
Cyclethon sparked the idea, but it could extend to walks, trips, and other IRL streams. What interested me was giving people that option—to share more of the experience together, wherever they happen to be.
What it does
The working demo combines a rider overlay on real prerecorded cycling footage with a guest's navigable simulated 3D bicycle POV. Guests pedal, steer, coast and follow independently. WebSockets share rider state; WebRTC carries the host return video and two-way voice, with a Web Audio mix-minus graph. A local compositor produces the host output for capture.
The demo labels prerecorded real footage and the simulated guest view. Its narration and music are editorial audio. WebRTC microphone tests use disclosed synthetic input. The cycling excerpt is credited to CDawgVODs Cyclethon; source footage and derived imagery remain excluded from the private GitHub repository.
How we built it
Built with OpenAI Codex, JavaScript, WebSockets, WebRTC, Web Audio, Three.js and the HAMON CC0 human rig. The main build session is 01a19dd7-b88d-73d1-bb74-f9f0c08f14a5. Bounded CLI peers and a durable Windows foreman integrated and tested candidates while preserving the original working baseline.
The intended actual-world pipeline uses licensed Street View imagery, owned photos or video as reconstruction inputs to Wildside/HAMON geometry. Guests navigate generated geometry. Prepared-area APIs track provenance, READY tile bounds, observed/inferred/unknown coverage and hash-verified loading. High-fidelity whole-area reconstruction and measured outdoor localization remain development work.
Validation and challenges
The verified private source release 5d26a9db1efd3db36aff227afc41cc450d70d02a is deployed on Raider and passes all 72 automated tests. It adds independent rider position and heading, corrected host follow distance, shared outfits and streamer presence. Fresh Windows Edge loopback on this exact release verified two guests, steering and coast, decoded host video and increasing two-way generated audio RTP with zero page exceptions. The replacement narrated demo uses fresh captures of this exact frozen release: moving host footage with the shared guest overlay, a separate 1080p guest capture, and actual public Twitch playback. It includes editorial narration and original music; microphone test tones are omitted. Earlier exact-pin Windows loopback QA verified two independent guests, control acknowledgments, decoded host video and bidirectional audio RTP. Those results are local test evidence, not physical-device or WAN certification.
The main challenge was keeping guest control, the host overlay, voice and capture coherent while making source quality explicit. Prerecorded footage and simulation let us test the complete interaction without pretending to have solved live outdoor localization. A post-submission capture verified the frozen 5d26 release broadcasting to TornadoZW on Twitch. The replacement video includes independently recorded public Twitch playback with the guest visible over moving footage. Hackathon registration and submission are tracked separately.
Built With
- hamon
- javascript
- openai-codex
- webrtc
- websockets
