Watch and listen
MediaPipe analyzes facial landmarks at 15+ FPS while a native local microphone stream captures your practice.
LOCAL • PRIVATE • OPEN SOURCE
Practice the talk.
See the evidence.
A presentation coach that measures camera contact, movement, pace, transcript gaps, and filler words—then ties every coaching action to an exact number and moment.
Runs on your Apple-silicon Mac. The interactive coach is never hosted in the cloud.
● FACE TRACKED15.1 analysis FPS
CAMERA CONTACT82%PACE138WPMFILLERS2
ONE LOCAL PIPELINE
Measurements first.
Language model second.
Build the evidence
whisper.cpp creates word timestamps. Python calculates pace windows, filler clusters, gaze breaks, movement, and quality flags.
Verify every claim
Python creates the allowed findings and actions. Ollama can prioritize them, but cannot invent a metric, value, timestamp, or judgment.
TRUST GATE
Direct coaching. Measured evidence.
A valid 30+ second recording receives actionable feedback using disclosed general practice bands. One reviewed baseline and two matched repeats add personal-reference comparisons.
- 1See strengths and specific practice actions immediately
- 2Review a baseline and capture two matched repeats
- 3Add verified personal comparisons
REPRODUCIBLE TEST EVIDENCE
Real videos.
Visible results.
12/12frozen tracking cases3/3licensed video pipeline tests30/30local LLM guardrail cases
The frozen anonymous face-tracking benchmark combines three regression cases, four development cases, four holdout-designated cases, and one in-memory derived two-face case. It adds eight Commons presentations/interviews to the three existing clips. Every case ran twice with an exact match across rounded deterministic aggregate tracking metrics; 12/12 passed. This is evaluation of a pinned MediaPipe model and local tracking logic—not identity recognition, face enrollment, or deep-model retraining. Media stays out of the repository and retains its Commons source license and attribution requirements.
Short, distant-face clip
14.9 FPS • 0% usable face presence • correctly marked face metrics insufficient.
Watch source + license ↗Mixed-shot clip
15.0 FPS • 36.7% face presence • 14 timestamped words • all quality flags emitted.
Watch source + license ↗Stable camera-facing address
15.0 FPS • 92.8% face presence • 602 timestamped words • all five inputs usable.
Watch source + license ↗Anonymous tracking benchmark
Checks exact inputs, 15 FPS cadence, dropouts, reacquisition, stable-segment landmark jitter, difficult-face abstention, and multi-face abstention.
Inspect the frozen manifest ↗PRIVATE BY ARCHITECTURE
Your practice stays yours.
No cloud inference
Camera, audio, Whisper, metrics, and Ollama execute on your Mac.
No raw-media archive
Raw audio is discarded after transcription. Camera frames are analyzed in memory.
Encrypted history
Session records use authenticated encryption and a random key stored in macOS Keychain.
Adversarially evaluated
30/30 synthetic cases passed, including missing data, excellent sessions, short recordings, and appearance bait.
MIT LICENSED
Run it. Inspect it. Improve it.
The local model recipe, backend, frontend, installer, tests, and full evaluation report are public.
Open the GitHub repository