Patchline bench

Live app · Source repository

Patchline Bench

Built by Rishik Rontala. A nonclinical sensing and signal-review prototype, kept separate from prior Patchline hackathon editions.

Run npm start, then open http://127.0.0.1:8790. Generate a reproducible steady, coordinated-shift or sensor-fault window. Inspect four charts and local model likelihoods; review an eligible synthetic window and log a virtual rehearsal. Import JSON locally or explicitly choose a USB serial port in a compatible browser. Bench and imported windows are always held. No hardware actuation or therapeutic dose exists.

The local classifier was trained from scratch on 1,200 simulated runs, selected using 300 validation runs, and evaluated on 300 independent test runs. Accuracy 89.7%, macro F1 89.4%. It missed 31 of 100 injected sensor faults; these metrics do not establish infection detection, medical benefit, or generalization to real sensors. A separate quality gate flags missing/held signals and prevents rehearsal for estimated faults or low confidence. Its review threshold is an explicit demo policy, not a clinical threshold.

npm test runs 35 independent software tests. python3 tools/board.py produces the passive PCB draft, netlist and drawing with layer-aware connectivity checks. node tools/generate-data.mjs and python3 tools/train.py reproduce the simulation/model. Both firmware modes pass native host-stub checks, including I²C failures. Both firmware modes compile for the selected ESP32; hardware measurements remain pending.

Read LIMITATIONS.md, hardware/WIRING.md and firmware/README.md before touching hardware. The board has no input protection or patient-contact design and still requires KiCad DRC and electrical review.

Featherless GLM 5.2 supplied an initial simulator implementation; it was corrected and tested locally. Featherless DeepSeek-V3.2 supplied an independent firmware review; verified findings and rejected errors are recorded under proof/. Secrets stay outside this project. The shared cumulative Featherless cap is $10; current conservative ledger estimate approximately $0.5811. Adaption is limited to 100 existing credits, with no purchases. Its proposed job fine-tunes a bench evidence summarizer; the numerical model above is already trained locally. Cloud training status is tracked separately and must not be represented as complete before confirmation.

The exploded assembly uses original vector illustrations, with real hub routing underneath. Click Assemble patch to close the six open concept layers, then Open layers to reverse it. The same button supports keyboard activation; reduced motion switches views immediately. Scrolling does not control the animation. A standalone version is packaged at artifacts/Patchline-Bench.html; direct browser testing of that file was blocked by the browser file-URL policy, while the source app was verified through its normal local preview.

Devpost draft

The local story is in devpost-submission.md. It separates working software from proposed business features, unassembled hardware, and the pending Adaption training experiment. Final event entry and the public five-minute voiceover video remain separate steps.

Demo videos and voiceover scripts

• 2:30 silent pitch demo · matching script
• 5:00 silent full walkthrough · matching script

Both videos combine actual browser captures with source/evidence cards. They are edited walkthroughs, not physical hardware footage. They contain no audio so Rishik can record his own narration. Exact duration, silent streams and full decoding were verified in proof/demo-verification.json. Record each timed chapter, place the voiceover over the corresponding MP4, and export before uploading a public/unlisted submission video. The five-minute version covers the Next Founders rubric; the shorter cut is for pitches and other submission limits.