Solum

Landing Page
Scanned Product
Top Recommendation
Chat Feature
Food Combinations to Avoid
Health Goals
Dietary Restrictions
AI Assisted Goals
User Settings

Inspiration

Most nutrition apps only track calories, protein, carbs, and fat. That leaves out anyone managing blood sugar, following a religious or cultural diet, or just trying to eat more fiber and produce. Comparing dense nutrition labels against a goal that isn't "lose weight" is still entirely manual, standing in the aisle. Solum puts that goal-specific reasoning directly into that moment.

What it does

Goals across four axes: health focus, dietary pattern, hard restrictions, free-form nutrient targets, not just calories and macros.
Scans a barcode, a label photo, or a whole shelf in one shot.
Falls back to Claude's vision when a barcode has no usable data.
Scores every product against your full goal profile in one call, with a plain-language reason and interaction warnings.
Answers nutrition questions by text or voice, with real tables, not paragraphs.

How we built it

Frontend: Next.js (App Router), TypeScript, Tailwind. Mobile-first PWA.
Backend: FastAPI (Python), pytest-gated deploys.
Data and auth: Supabase Postgres and Auth, row-level security.
AI: Claude (Anthropic API, AWS Bedrock fallback) for vision, scoring, and chat.
Voice and identity: ElevenLabs, Persona.
Deploy: GitHub Actions, OIDC, AWS ECS. No stored secrets.

Scan pipeline: $$\text{Barcode} \rightarrow \text{Open Food Facts}$$ $$\text{No match} \rightarrow \text{Claude Vision} \rightarrow \text{Structured Nutrition Data}$$

Recommendation pipeline: $$\text{Goals} + \text{Scanned Products} \rightarrow \text{Claude (one call)} \rightarrow \text{Scored Picks} + \text{Warnings}$$

Chat pipeline: $$\text{Question (text or voice)} \rightarrow \text{Claude} \rightarrow \text{Text + Table Blocks} \rightarrow \text{ElevenLabs (optional)}$$

Access gate: $$\text{3 Free Calls} \rightarrow \text{Persona Check} \rightarrow \text{Unlimited Recommendations}$$

Challenges we ran into

Redesigned goals mid-build: one enum to four flexible axes, no migration needed for new nutrients.
Made Claude's output renderable: a typed schema of text and table blocks, not parsed prose.
Real camera UX: viewfinder framing that worked in a mock broke on a real phone.

Accomplishments that we're proud of

One recommendation engine, full goal profile, one Claude call, self-explained.
One scanning pipeline for three real inputs: barcode, label, shelf.
Shipped to real infrastructure: containerized, zero stored credentials, not a laptop demo.