← Lyon Works

Eat This Lor 吃這咯

The food recommender that just decides for you.

01 — The Spark

Decision fatigue is real, even for food.

Eat This Lor solves the one food problem apps keep ignoring — not discovery, but decision.

It's not that we don't know what food is out there. We all have our hawker centre regulars, our go-to mamak stalls. The problem is the daily paralysis of choosing from what we already know. Existing food apps show you 500 options. Eat This Lor just tells you what to eat.

Who it's for

The spouse who can't decide 👩

Always asks "what to eat?", rejects every suggestion, has a mental list but still can't commit.

Who it's for

The office colleague on repeat 💼

Asks daily but won't walk far. Defaults to the same chicken rice stall for the 4th time this week.

Who it's for

The team lunch planner 👥

Either gets creative and burdens the group, or everyone defaults to the lazy option. Lunch at 2pm.

02 — The Dig

We surveyed 17 Singaporeans about how they actually think about food.

Before building, we ran a survey testing food similarity perceptions and eating habits. First we established context: food decisions are a daily recurring problem.

59%

eat out 5–7 times a week

29%

eat out 8–10 times a week

12%

eat out more than 10 times a week

For most people, choosing what to eat isn't an occasional decision — it's a daily friction point.

💡

The insight we didn't expect

We assumed cuisine origin (Chinese, Malay, Western) would drive food similarity. The data said otherwise. When asked what factor matters most, 35% said main ingredient — while only 24% said cultural origin. The similarity scores made it even clearer: Mala Tang and Korean Army Stew, two different cuisines but both heavy spicy soups, scored the highest for similarity among all cross-cuisine pairs (2.88/5). Vietnamese Pho and Ban Mian — different cuisines again, both light noodle soups — followed closely (2.59/5). Cooking style crosses cuisine. That was the finding.

This flipped our entire recommendation logic. Instead of grouping by cuisine, we group by Staple Base first, then Style, then cuisine last. The weights reflect what people actually feel — not what menu categories say.

17

survey respondents

14

dish pairs tested

20

dishes categorised

73

dishes in database

03 — The Build

Two modes, one goal: no more cannot decide.

Spin the roulette and let fate pick, or answer 3 quick questions for a tailored suggestion. Either way, you get an answer — not 500 options.

Live Demo

How hungry are you right now?

Spicy or not?

What kind of base are you feeling?

🍽️
Finding your match...

04 — Under the Hood

Rules-based logic, research-driven weights.

The recommendation logic is intentionally rules-based — fast, predictable, and transparent. Each of the 73 dishes is tagged across three dimensions, weighted by what the survey showed actually drives food decisions.

🍚 Level 1 — Staple Base 50%
Rice-based Noodle-based Bread-based Protein-based Soup-based Others
🍲 Level 2 — Style 25%
Heavy / Light Spicy / Mild / Non-spicy Soup / Dry / Wet / Cold
🌏 Level 3 — Cuisine Origin 15%
Chinese Malay/Peranakan Indian Japanese Korean Western Vietnamese Thai

Cuisine carries the lowest weight because the survey showed it's the least predictive of whether you'll enjoy a recommendation. Your gut doesn't care if it's Chinese or Japanese. It cares if it's soupy and heavy.

05 — What I Learned

The research changes everything. The prototype proved it.

The original vision was bigger: a learning model that tracks what gets recommended, what you actually eat, and uses that feedback to get sharper over time. Pair that with location data and live F&B menus — and you have something genuinely useful at scale.

What held it back was maintenance, not ambition. A live service needs tending. That's a different kind of commitment. What this project proved instead: rapid prototyping with AI has matured to the point where builders can test ideas in hours, not months. The barrier to starting has dropped.

✓ What Worked

Doing real user research before touching any code — and it completely overturned the core hypothesis before a single line was written. The prototype came together fast. And proving that AI-assisted building can compress months of effort into hours is itself a finding worth sharing.

↯ The Constraint

The gap between prototype and product is real. Mapping enough food items to be meaningful, then layering location data on top — both are solvable, neither is trivial. For now, the prototype does exactly what it promises. You'll get a response that says: eat this lor.