BrosBHustlin – Finance Bro Energy Detector

Inspiration

We were inspired by the humor and absurdity of finance bro culture — Patagonia vests, luxury watches, AirPods, and endless corporate buzzwords.
We wanted to create a fun, interactive way to measure and gamify “bro energy” while combining AI, satire, and gameplay into a single web app.

What it does

BrosBHustlin is a satirical web app that:

Detects finance bro items via a custom-trained YOLO model.
Analyzes speech for corporate buzzwords in 30-second clips.
Calculates a total score out of 400 points and assigns tiers:
- Peasant 🥔
- Analyst
- Associate
- VP of Cringe
- CEO of Insufferable 👑
Turns your score into a sperm race game, where higher points give faster, more competitive races.

Scoring Formulas:
$$ \text{Item_Score} = \sum_{i:\ \text{confidence}_i \ge 0.3} (\text{confidence}_i \cdot \text{points}_i) $$ $$ \text{Voice_Score} = \min(200, \text{buzzword_count} \cdot 25) $$ $$ \text{Total_Score} = \min(400, \text{Item_Score} + \text{Voice_Score}) $$ $$ \text{Tier} = \left\lfloor \frac{\text{Total_Score}}{80} \right\rfloor $$

How we built it

Computer Vision: Trained YOLOv8n from scratch on CPU using 300+ labeled images. Every clothing item was manually annotated; pre-trained models and paid vision APIs were not used.
Speech Analysis: Whisper ASR transcribed audio and flagged buzzwords using regex.
Web Stack: React + Tailwind for frontend, Flask backend, SQLite leaderboard.
Gamification: Score drives a sperm race game, creating a competitive, interactive layer on top of the AI scoring.

Challenges we ran into

Data Annotation: Labeling every single clothing item was time-consuming.
No Pre-Trained Models: Training YOLOv8 from scratch on CPU slowed experimentation.
Lighting Variability: Model accuracy decreased in inconsistent lighting.
Resource Constraints: CPU-only inference required careful optimization of image resolution, batch size, and augmentations.

Accomplishments that we're proud of

Successfully trained a custom object detection model from scratch with 91.2% mAP50.
Integrated multi-modal AI (vision + NLP) in a web app.
Created a fun, shareable gamification mechanic (sperm race) tied to AI scoring.
Built the full stack with React, Flask, and SQLite, achieving real-time interactions on CPU.

What we learned

Training from scratch is feasible but requires careful dataset prep and augmentation.
Multi-modal AI can create rich and interactive experiences.
Gamification can turn raw AI outputs into engaging, viral-ready gameplay.
Optimization for CPU inference taught us resource-efficient deep learning techniques.