This is a submission for the DEV x Presage SmartSpectra SDK Challenge.
TL;DR: Game designers draw the tension they planned for a level. Senti watches the playtester through a normal laptop webcam (Presage SmartSpectra) and shows where they actually flinched, spoke or looked away. Every gap between plan and player becomes one clickable sentence that jumps the recording to that moment.
What I Built
"I built a jump scare. Did anyone actually jump?"
Senti is a Windows playtest tool for indie game designers. You draw the tension curve you planned for a level (calm, rising, jump scare, boss) and drop named beats on it. The playtester agrees on a consent screen and plays. A normal laptop webcam, read by Presage SmartSpectra, notices what they actually did: startle head-jerks (I call them flinches), voice reactions, looking away.
The report puts the felt curve over your intended curve and the gameplay recording. It flags every mismatch in a plain sentence, like "0:50 to 1:10: you planned a peak, the player barely moved." Click the sentence and the video jumps to that moment.
Players are polite in surveys, and facecam footage is endless. Senti starts from the designer's own plan and shows where the game and the player disagreed, second by second.
Demo
Try it: download Senti-Setup-0.1.0.exe from Releases, paste your free Presage key, click New playtest. No game handy? The repo has an original 2-minute test level in demo-level/.
Code
Techkeyy
/
senti
Senti: draw how your game level should feel, playtest it, and see where players actually reacted. Webcam reactions via Presage SmartSpectra.
Senti
Draw the tension you planned. Watch the tension your player actually felt. A Windows playtest tool for indie game designers, read live from the webcam with Presage SmartSpectra.
Watch the 2.5-minute demo · Download · Evidence: real session data · Report screenshot
"I built a jump scare. Did anyone actually jump?"
You plan how a level should feel: calm, then tense, then a scare, then a boss. Then you hand it to a playtester and ask what they thought. They are polite. You end up scrubbing through hours of facecam footage, guessing at the moments that mattered.
Senti is for the designer who runs those playtests. You draw the intended tension curve with named beats, the playtester plays, and the webcam notices what they did: startle head-jerks (I call them flinches), voice reactions, looking away, and a little expression data. The report lays the felt curve over your intended…
Electron + React + TypeScript, with @smartspectra/node-sdk 3.5.0 running the native camera in a utility process. The felt model (src/shared/felt.ts) is pure, deterministic and unit tested.
How It Responds to Human Signals
I tried the obvious route first: build the curve from facial expressions. On my face and lighting, the expression scores sat at about 60% "surprise" at rest and did not separate calm from tense. So I wrote a test with the pass rule fixed in advance, watched it fail, and changed the approach.
Senti reads behaviour, measured against each player's own baseline:
- Head speed. SmartSpectra gives 478 face landmarks about 30 times a second. Senti tracks the face centre's speed in face-widths per second. A 30 second resting baseline learns how still this person normally is. A flinch is a second whose peak speed is more than twice their baseline 95th percentile.
- Voice. SmartSpectra's talking detection marks gasps, laughs and "nope".
- Attention. If the face is lost for 2 seconds or more, that is reported as looking away. It never counts as "calm".
Each flinch adds an impulse that decays over 4 seconds, and that builds the felt curve. Mismatches are checked in windows around each beat and in 10 second tiles elsewhere. Every verdict comes from simple rules, with no LLM and no cloud call of my own.
Evidence (all real, all in the repo):
- Pre-registered startle test: 11 spike events during jump scares vs 1 in calm. The recorded trace is asserted by a unit test.
- Fresh-install playtest, 18 Nov: the player stayed flat through a planned rise from 0:50, then flinched hard at 1:13 and turned away for 2 seconds. Senti flagged all three gaps.
- Owner playtest: face tracked 136 of 136 seconds at 29.8 readings per second, 0 restarts, on a 2017 dual-core laptop.
The two session exports are in docs/evidence/ (no camera frames, no video), so anyone can check the numbers.
Building on SmartSpectra, things I hit:
-
reset()can block the event loop, so I destroy and rebuild instead. - A
videoOutputlistener cut processing from about 28 to 5 fps on two cores, so I attach it only during the camera check. - The SDK stops itself below about 25 fps. Running the worker at high OS priority held 30 fps while a game played and the screen recorded.
There are more notes in the README.
Prize Categories
- Most Creative Use of SmartSpectra: a webcam signal pointed at a new user. It serves the game designer, not the player or a streamer, and lines up their intent against the player's behaviour.
- Best Thoughtful Real-World Use: a behaviour-only design that is honest about what a webcam can see. It gets consent first, keeps everything local and never saves camera frames.
Consent and Data Handling
- Camera: used only after the playtester agrees on the consent screen. Only one person should be in view.
- Camera frames are never saved or uploaded. The only picture of the face is the live preview on the camera-check screen. During play, only the screen is recorded, never the face.
-
Where data lives:
%APPDATA%\senti\sessions\<id>\holdssession.json(intended curve, beats, per-second readings, baseline) andgameplay.webm. Senti sends nothing anywhere. - Deletion: the trash icon on Home deletes the whole session folder after a confirm.
-
API key: encrypted with Windows via Electron
safeStorage, never logged and never shown back. -
Telemetry: the SDK's own telemetry is off (
enableTelemetry: false). - What Presage sees: the SDK contacts Presage to check the key and usage. Per Presage's documentation, it may also send short derived number summaries (never video) for an LLM wellness feature that apps cannot switch off. Senti does not request, show or store those insights.
- Not a medical device. Senti makes no health or emotion claims. It reports head movement, talking and whether a face is visible.
Honest limits: Windows x64 only. Tested by one person on one laptop. The thresholds were tuned on one person and have not been validated across players, skin tones or cameras.



Top comments (0)