What I Learned Building a Full-Stack AI App with AWS + Vercel in 7 Days
Build retrospective of ChemSpectra Agent — an AI-powered FTIR spectral analysis app built with Next.js on Vercel, FastAPI, and Amazon DynamoDB for the H0 AWS+Vercel Hackathon. Covers lessons on v0 prototyping, DynamoDB GSI design, and ReAct agent reasoning.
Reflections from the #H0Hackathon — building ChemSpectra Agent
Last week I shipped a full-stack AI app for the H0 hackathon: Next.js on Vercel, FastAPI backend with ReAct reasoning, Amazon DynamoDB for persistence. It analyzes FTIR spectra for materials QC labs — a B2B use case with users in 52 countries.
Here is what I learned.
Vercel v0: Real Prototyping Power
v0 generated a working Next.js frontend in two hours — file upload, analysis cards, confidence badges, dark theme. Clean TypeScript and Tailwind. I still wired up the API client, SSE streaming, chat interface, and history sidebar, but v0 eliminated the blank-canvas problem.
Takeaway: Use v0 for the 80% that is standard UI. Spend your time on the 20% that differentiates.
DynamoDB: Simple Until You Query
Setup was fast: table, partition key, TTL, done. Single-item reads by
session_id are instant. The friction came when listing recent sessions — a naive scan returns random order. The fix: a GSI (gsi-created) with a fixed partition key and created_at sort key, turning O(n) scan into O(1) query. I added a second GSI (gsi-material) for "show all analyses of Polyethylene" aggregation, atomic counters on a stats table for usage metrics, and conditional writes to prevent overwriting confirmed sessions.Takeaway: Design your access patterns before writing code. Plan GSIs upfront — retrofitting them costs 5-10 minutes of CREATING status per index.
ReAct Loops Work
I was skeptical about letting the LLM choose tools autonomously. In practice, Qwen-3.7-Max consistently picked sensible combinations — most analyses completed in 2-3 iterations, never hitting the 6-round cap. The self-verification (confidence < 0.75 triggers re-investigation) was the most satisfying feature. Watching the agent detect contradictions and call additional tools to resolve them felt like genuine reasoning.
Infrastructure Eats Time
Seven days broke down to: Day 1-2 frontend + Vercel, Day 2-3 AWS setup, Day 4-5 integration + CORS + SSH tunnels for public access, Day 5-7 video + docs + polish. Infrastructure consumed more time than features.
Five Tips for Hackathon Builders
- Deploy first. Get Hello World on your production stack before writing features.
- Use code generation aggressively. v0, Cursor, Claude — whatever kills boilerplate fastest.
- Match database to access pattern. DynamoDB for key-value sessions; would be wrong for relational queries.
- Script your demo video first. It drives feature prioritization.
- Ship something real. Judges know the difference between demos and products.
Building under deadline, with real constraints, for a real use case — that is when the best decisions happen.
This article was written as part of my participation in the H0 AWS+Vercel Hackathon.
*Tags: #H0Hackathon #AWS #Vercel #DynamoDB #AI #Hackathon #FullStack #NextJS #FastAPI #BuildInPublic
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article