Status: private (handles personal health data; code shown) · Code: 02_ai-engineering-code/blood-suga/ (src-ai/groq-vision.ts, src-ai/meal-photo-analyze-route.ts, eval/eval-meal-photo.ts, eval/dataset-metadata-SAMPLE.csv)
Problem. A glucose-management dashboard needs to estimate carbs and calories from a photo of a meal — accurately enough to be useful, safely enough not to give medical dosing advice.
What I built.
dish_id, total_mass, total_calories, total_carb, total_protein, ingredients_json, plus glycemic-index/load supplements). eval-meal-photo.ts runs accuracy regression against the labels — MAPE, median % error, in-range hit rate for carbs and calories.Why it matters for an AI-native role. This is the full applied-AI loop most candidates can't show: production vision-model integration with structured output + schema enforcement + safety screening, plus dataset curation and a quantified eval harness. "I don't just call the model, I measure whether it's right against labeled data" is exactly the muscle these companies hire for.
This one shows the full loop. The product estimates carbs and calories from a meal photo using a Groq vision model, and the output is a structured JSON range validated against a Zod schema, with safety screening so it never gives insulin dosing advice. But the part I care about is the eval: I curated a fifty-eight-hundred-image nutrition dataset with a five-thousand-row metadata file, and I wrote a harness that scores the model against those labels, MAPE, median percent error, in-range hit rate, for both carbs and calories. So I can actually tell you how accurate it is and catch regressions. That's the difference between calling an API and doing applied AI.
GG · Applied AI Engineering portfolio · synced to youdontneedmy.help design