A Shishito Pepper Analyzer

How to Tame the Green Hot Chili Peppers

Introduction

You’re sitting at a restaurant and they bring out the first course. Ah, some scrumptious shishito peppers. You bite into the first one: delicious. You bite into the second: delic- oh wait, IT BURNS IT BURNS IT BURNS.

The vast majority of shishito peppers are mild, but about one in every ten are significantly hotter. What if there was a way to detect which shishito peppers were the hot ones?

The Problem

The spice level of a given shishito pepper is based on the amount of capsaicin, a chemical compound, produced during growing. While shishito peppers are bred to be mild, about one in ten develop unusually high capsaicin levels, the result of random genetic variation and environmental stressors like inconsistent watering or temperature swings.

At present, there is no way to distinguish between mild and hot shishito peppers based on their appearance at the dinner table. Some may seek the thrill of the surprise hot pepper, but for those who want to enjoy the taste without fearing the heat, it would be useful to have a way of discerning which peppers are spicy.

The Idea: a shishito pepper analyzer that can distinguish hot from mild shishito peppers.

Option 1: Machine Learning + Video Analysis

How it would work: you’d take a picture of the peppers on your phone and the model would predict which peppers are hot. Think Jian-Yang’s SeeFood app from Silicon Valley (except instead of telling you whether the food in question is a hot dog or not, it tells you if it’s spicy). This would require training a model on many shishito peppers to be able to predict spice levels based on visual differences between spicy and mild shishito peppers.

Challenge: there are no clear visual / physical differences between mild and spicy shishito peppers (if they were obvious, then a tool would be unnecessary). Even if subtle differences exist, the model would require lots of training data and extensive labeling to reliably distinguish.

Option 2: Spectroscopy

How it would work: spectroscopy is a technique that analyzes how a material reflects or absorbs light. Capsaicin absorbs light at known frequencies, so a tool could be created to detect relative capsaicin levels in different peppers.

I asked ChatGPT for the absolute cheapest way of building this. It involves buying a bare near infrared (NIR) spectrometer module ($60), pairing it with an Arduino ($5), creating a light setup with NIR LEDs ($10), testing 50 peppers and manually logging which ones are spicy ($10), and then using free tools in Python to train a model that will correlate the spectral patterns from the spectrometer with spice levels.

This is the absolute cheapest way of doing it, sacrificing complete accuracy for cost-efficiency, and it still runs over $80. While it would be an impressive technical feat to create a functioning prototype on such a small budget, it’s likely too pricey for the average shishito pepper enjoyer.

The Real Use Case

The most realistic builder of this technology is not an individual consumer but restaurants. Imagine a restaurant offering you the ability to custom-design your shishito pepper appetizer based on exactly how many spicy peppers you want in it. The spectrometer could pay for itself in the volume it would drive (or worst case the restaurant slightly raises the price of shishito peppers).

Logistically, the restaurant would use the spectrometer to scan each pepper in a given batch of peppers and then mix and match to meet the customers’ request. This shouldn’t be too time-intensive since a serving of shishito peppers is usually no more than ten peppers and the spectrometer doesn’t take more than a couple seconds per pepper.

The restaurant could even get more mileage by applying the device to other spicy menu items, although this would require training the model on each new food. Since the largest fixed cost is the hardware, the restaurant can achieve economies of scale with every additional use.

The restaurant would likely need to invest in additional safety, cleaning, and technological features to make it viable, with ChatGPT estimating the minimum cost of a restaurant-ready version at $350. So still pretty expensive and unlikely to be adopted absent a strong consumer movement in favor.

Conclusion

I would love to see someone give this a try, at the very least as a publicity stunt. In the meantime, I’ll continue rolling the dice.