PocketGym — Final Presentation Reference

01 Intro Kobe · 25s

Greet audience, introduce team: Aakriti, Iggy, Evan, Kobe
"PocketGym — a web app that uses your webcam to count reps and grade your form, in real time, with no special hardware"
Live: pocketgym.evanmusick.dev

Aakriti — what's the problem?

02 The problem Aakriti · 35s

Here's what we built.

03 What we built Aakriti · 45s

Evan will explain how this is wired together.

04 Architecture Evan · 60s

Two tiers
Browser tier does all CV: webcam → MediaPipe Pose (WebAssembly) → 33 landmarks → angles → classifier → form scorer → overlay
All ML happens in the user's tab. Video frames never touch our server. Privacy by default; cheap hosting
Bridge: POST /api/sessions sends a small JSON summary at session end
Server tier: Flask 3.0 + Flask-Login + SQLAlchemy → Postgres (SQLite in dev)

Kobe — how does the form scoring layer work?

05 Form scoring Kobe · 60s

Rule-based — no model training, no labeled dataset
Four steps every frame: pose → joint angles → per-joint score → weighted average
Joint angle: θ = (180/π)·arccos((A−B)·(C−B)/(‖A−B‖·‖C−B‖))
Per-joint score: s = max(0, 100 − 2.5·|measured − ideal|)
Ideal ranges tuned by hand against test cases
Trade-off: hand-tuned rules vs needing thousands of labeled videos

Evan — give the audience the tech stack.

06 Tech stack Evan · 35s

Browser: MediaPipe Tasks Vision (pose_landmarker_full.task), vanilla JS, server-rendered SVG charts via Jinja, Web Speech API
Server: Flask 3.0, Flask-Login + bcrypt, SQLAlchemy, SQLite dev / Postgres prod
Deploy: Render.com behind Cloudflare DNS + TLS
No build step, no React — boring on purpose

Iggy — walk us through the sprints.

07 Sprint progression Iggy · 50s

Sprint 1 (Mar 6–26): core pipeline — pose, angles, classifier, form scorer; Python prototype against a webcam
Sprint 2 (Mar 27–Apr 9): users + data — accounts, login, profiles, history, per-set form charts
Sprint 3 (Apr 10–28): the big shift — ported entire pipeline to JS for browser, deployed to Render, Cloudflare DNS, UAT on live URL
That's what's deployed today

Now we'll show it running.

08 Live demo ALL · ~10 min

Each presenter drives their own segment — see the demo segments table below
Pre-flight: tab open at /auth/login, webcam tested, backup recording in second tab, volume up
If anything fails: switch to backup recording, narrate, return to slides. Don't apologize, don't troubleshoot live

Here's what this whole project taught us.

09 What we learned Iggy · 60s

Browser is a real ML runtime: WebAssembly + MediaPipe = 30+ fps pose inference, no GPU, no install
Same algorithm in two languages drifts: Python prototype + JS production diverged on tweaks; needed shared test fixtures
A live URL finds bugs localhost never will: TLS-only camera permission, mobile viewport quirks, voice API differences
How a project introduces itself matters: README + landing page rewrites changed how new users + graders engaged

Aakriti — who built what?

10 Team Aakriti · 35s

Read the four cards aloud
Evan: scaffold, pose pipeline (PE-01..03), user prefs, production deploy, pose-math test suite
Aakriti: rep counting (RT-10), set tracking (RT-11), accounts (UM-12), profiles (UM-13), progress charts (DA-21)
Iggy: joint angles (PE-04), classifier (ER-05), additional exercises (ER-06), history log (DA-15), audio coach (FA-18)
Kobe: ideal angles (FA-07), form scoring (FA-08), correction tips (FA-09), post-workout summary (DA-16)

Evan — what would we do differently?

11 If we did it over Evan · 50s

Browser-first from day one — Sprint 1+2 of Python prototype cost us time we then spent porting in Sprint 3
Test Python and JS versions against the same fixtures — caught the drift late; a shared video-in/scores-out fixture would have caught it in Sprint 2
Stand up the deploy in Sprint 1 — Render, DNS, TLS, mobile bugs only show up live; treat the deploy as a feature
Build a labeled dataset for the classifier — rule-based works, but a trained model could be more accurate and easier to extend

Kobe — close us out.

12 Conclusion Kobe · 2–3 min

Recap in one breath: webcam-only, real-time form coaching, privacy-by-default, 5 exercises live in production
Honest about scope: rule-based scoring (not ML), 5 exercises (not 50)
Live at pocketgym.evanmusick.dev · source at github.com/musickevan1/pocket-gym
Thank the audience, open for questions