Skip to content

rateradar

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

rateradar

README.md

RateRadar

Compare live loan rates and savings yields across Vietnam's banks, with Shinhan Bank featured — then auto-fill applications at every bank in parallel, stopping just short of submitting.

Built for the Financial Services track of Agentic AI Build Week (aabw.genaifund.ai), sponsored by Shinhan Bank / Shinhan Future's Lab Vietnam. **Live Link - ** https://rateradar-teal.vercel.app/

A framing note

Financial Services is officially an enterprise track (the end user is meant to be a business team, not an individual). This build is intentionally consumer-facing instead, because that's what the literal brief describes: a rate comparison tool with parallel auto-apply. The honest defense for judges: this is Shinhan's customer-acquisition and rate-transparency play — Shinhan is the featured, sponsor-anchored bank in every comparison, and the tool exists to capture a shopper's attention and data at the exact moment they're rate-shopping.

What it does

  1. Pick a product — a 12-month savings deposit, a home loan, or a personal loan
  2. See live rates across Shinhan Bank and 5 real competitors (Vietcombank, Techcombank, VPBank, BIDV, MB Bank), sorted best-first — savings compares highest APY, loans compare lowest interest rate
  3. Enter your details once — name, amount, tenure
  4. Auto-fill applications at every bank in parallel — one real-time TinyFish agent per bank, filling that bank's actual application/inquiry form with your details
  5. Review and submit yourself — the agent stops before the final submit step on every bank's site. It never completes a binding financial application on your behalf; you get a direct link to each bank's form plus an estimated monthly payment/interest, and you finish it there

Shinhan's own rates come from an internal feed, never scraped — a bank already knows its own published rates.

Architecture

.github/workflows/sweep.yml
  Daily cron, 07:00 UTC = 14:00 Vietnam time
        |
        v
scripts/sweep.ts  -->  POST /api/sweep (deployed app)
        |
        v
lib/sweep.ts: runSweep()
  1. internal.ts   -> Shinhan's own rates, NEVER scraped
  2. rates.ts (x5) -> real-time TinyFish agents, one per bank, all 3
                      products read in a single pass each
        |
        v
lib/storage.ts -> Upstash Redis (persists across requests/deploys)
        |
        v
GET /api/state (read-only, never triggers a sweep in production)
        |
        v
Dashboard: rate comparison + applicant form
        |
        v
POST /api/apply -> one TinyFish agent per bank fills the real
                    application form, stops before submit
        |
        v
Staged applications, reviewed and submitted by the person themselves

Tech stack

Layer Technology
Framework Next.js 15 (App Router)
Language TypeScript
Rate + form-fill agents TinyFish Web Agent API (@tiny-fish/sdk, stealth profile, queue + poll)
Scheduling GitHub Actions (daily cron, 14:00 Vietnam time)
Persistence Upstash Redis (falls back to a local JSON file in dev)
Styling Tailwind CSS
Deployment Vercel

How to run locally

npm install
cp .env.example .env.local   # add TINYFISH_API_KEY (optional)
npm run dev

Open http://localhost:3000. First local run bootstraps one rate sweep automatically. No TINYFISH_API_KEY? Every bank falls back to realistic simulated rates, so the demo always works.

Environment variables

Variable Description
TINYFISH_API_KEY TinyFish agent key. Unset → rates and applications are simulated.
UPSTASH_REDIS_REST_URL / UPSTASH_REDIS_REST_TOKEN Free Redis at upstash.com.
CRON_SECRET Any random string, protects /api/sweep from random requests.

Deploying

  1. Push to GitHub, import into Vercel, add the four env vars above.
  2. On the GitHub repo, add Actions secrets APP_URL (your Vercel URL) and CRON_SECRET (same value).
  3. .github/workflows/sweep.yml fires automatically every day at 14:00 Vietnam time.

Notes for the pitch

  • The one hard safety boundary: the agent never submits a binding financial application. It fills every field it can find and stops — the person reviews and clicks submit themselves, bank by bank.
  • Shinhan is never scraped for its own data (it already knows its own rates) — only competitors are, which is the actual hard intelligence problem worth automating.
  • Sorting correctly handles that "best" means opposite things for the two product kinds: highest APY for savings, lowest rate for loans.