Nutrimetry — personal project, in production

Nutrition tracking with laboratory-grade precision.

A PWA I use every day: 108 micronutrients, phased targets, semantic search and AI analysis. Built end to end — product, data, design and infrastructure — by one person.

Try the live demo Read the code

Designed and built by Bryan Rodríguez Abarca · part of the vryahn.com portfolio.

Today · Jul 16 On target

1,842 / 2,100 kcal

Protein132 / 150 g

Carbs178 / 220 g

Fat58 / 70 g

Sodium · 1,240 mg ⚠ under 1,500 mg

Real interface data model — the safety floor never turns off.

108micronutrients

163automated tests

20SQL migrations

4.5:1minimum contrast

Decisions, not just features

Every trade-off resolves in favor of data accuracy. These are the calls that shaped the system.

Data integrity

Accuracy wins every trade-off

Everything is stored per 100 g and computed in SQL views at read time — values are never copied between tables. Doubtful data gets flagged ⚠, never saved silently.

Declared kcal are checked against Atwater factors (4/4/9, fiber at 2 kcal/g): a gap over 25% demands review before saving.

label · OFF · USDA · AI→ normalized to 100 g→ SQL view at read time

Banana	stored	logged: 150 g
kcal	89	134
protein	1.1 g	1.7 g
potassium	358 mg	537 mg
fiber	2.6 g	3.9 g

⚠ 89 kcal declared · 132 per Atwater — needs review

One correction to the food fixes every historical entry — the view scales, it doesn't copy.

Semantic search over the catalog

Hybrid retrieval: exact match answers first; if it comes up short, pgvector completes with cosine similarity over 768-dim embeddings generated when each food is saved.

The same infrastructure powers 'Ask your log': a 3-step RAG pipeline that answers questions citing only real logged data.

match_foods('yogurt')

food	cosine
Greek yogurt, plain	0.21	✓
Drinkable strawberry yogurt	0.28	✓
Cottage cheese	0.47	✓
Whole milk	0.68	✗ > 0.65

Spelling variants and synonyms stop breaking search — embeddings absorb them; a distance cap keeps noise out.

Security

Privacy by design

Row Level Security isolates every user inside Postgres: the filter lives in the database, not in the UI. No secret key ships to the client.

The repo is public with zero credentials in its history — the design assumes the code can be read by anyone.

create policy foods_select
  on nutri.foods for select
  using (owner = auth.uid());

session: user A Greek yogurt Oats Café de olla

user B's rows ████████████ ██████████ █████████████

Postgres doesn't even return the other user's rows — it's not a UI filter a bug could skip.

A deliberately small stack

~45 source files, zero TypeScript, no state library. Restraint is a feature: every dependency must justify itself in one line.

React PWAVite · Tailwind · Recharts · offline · es/en

supabase-js · row-level RLS

Postgres · SupabaseSQL views · pgvector · Auth · Storage

Photo or labelcanvas compression → 1024 px

response_schema JSON

AI cascadeGemini 3.8 → … → 2.5 → Groq → Mistral

MCP clientsClaude · ChatGPT · Perplexity…

MCP · OAuth

The same log8 tools: log, search, audit

git push mainGitHub Actions: lint · 186 tests · build

green CI = deploy

Vercelproduction — no staging, verified locally

ReactSupabase · Postgres + RLSpgvectorTailwindRechartsVite PWAGemini · Groq · MistralMCPGitHub ActionsVercel

Behind the project

nutri. is a personal project — precise data, honest interfaces, end-to-end delivery: from the SQL schema to the pixel.

Read the case study View portfolio GitHub

Nutrition tracking with laboratory-grade precision.

Decisions, not just features

Accuracy wins every trade-off

Write local, sync later

A fallback cascade, gated by evals

Semantic search over the catalog

Privacy by design

Its own MCP connector

A palette that never lies

A deliberately small stack

Behind the project