DebtClear

Crush your debt with real math, AI strategy, and a word-for-word phone script to settle each debt for less — built for people drowning in credit card and consumer debt.

Comment

Inspiration

Most debt payoff tools show you a chart and leave you alone with it. The hard part isn't calculating the avalanche — it's knowing what to say when you call your creditor, whether you even have leverage to negotiate, and what happens to your specific numbers when life changes. We built DebtClear to close that whole gap: real math, an AI advisor that has your full picture in front of it, and tools that turn the abstract plan into something you can actually execute.

What it does

DebtClear takes your debts (balance, APR, minimum payment) and monthly income, then:

  1. Simulates four scenarios — Avalanche, Snowball, minimum-only baseline, and a what-if slider you drag in real time to watch your payoff date and total interest update.
  2. Calculates a financial stress score (0–100) based on debt-to-income, payment burden, and weighted APR — and tone-matches every AI response to that severity tier (a healthy borrower hears "optimization", a distressed one hears "stabilization").
  3. Generates a personalized AI analysis — three paragraphs grounded in your exact numbers, no generic advice.
  4. Live AI chat advisor — ask anything follow-up ("what if I lose my job?", "should I refinance?", "is bankruptcy worth it?") and the AI answers with your actual debts, income, and stress score in front of it.
  5. Voice phone roleplay — practice the negotiation out loud against an AI creditor named Sarah Mitchell, with hidden settlement floors and a manager persona that pushes back. Browser speech recognition + synthesis for a real conversational feel; full text fallback when voice is blocked (use chrome browser for best results ).
  6. Word-for-word phone script — opening, hardship statement, initial offer, counters, closing — generated specifically for your leverage score.
  7. Settlement letter PDF — formal certified-mail-formatted letter generated by AI, downloadable as a polished jsPDF document.
  8. Crisis resources panel — when stress score ≥ 75, surfaces NFCC, CFPB, 211.org, and Upsolve. Empathy when the math is brutal.
  9. Cost of waiting card — pick 3 / 6 / 12 months delay and see the dollar cost of inaction at your weighted APR.
  10. Victory timeline — a horizontal celebration timeline showing exactly which month each debt gets eliminated, ending in a debt-free trophy.
  11. "Why your plan matters" comparison — side-by-side vs. minimum-only with a hero stat: "$X saved, Y years back, Zx less interest paid".
  12. Full plan PDF export — multi-page branded document (cover, strategy, advisor notes, prioritized debt list, action steps) for taking offline.

How we built it

  • Backend: Django 5 + Django REST Framework, pure API. Pure-Python simulation engine that models month-by-month interest accrual, minimum cascading, and priority-based extra-payment pooling. Separate engines for leverage analysis, settlement range, script generation, voice roleplay, settlement letter, financial advisor, and follow-up Q&A chat.
  • AI layer: Llama 3.3 70B (Groq) with a multi-key failover pool — primary key → secondary key → deterministic Python fallback. Fast-fail HTTP client (max_retries=0, hard timeouts) so a rate limit never stalls a user. Every AI endpoint has a working deterministic fallback so the demo never breaks.
  • Frontend: Next.js 14 App Router + TypeScript + Tailwind + shadcn/ui. Framer Motion for transitions, Recharts for the payoff timeline, jsPDF for both the settlement letter and full plan export. A WebGL fragment shader (cosmic nebula background) rendered at adaptive resolution — 35% on mobile (20fps), 60% on desktop (30fps).
  • Voice features: Browser-native Web Speech API — SpeechRecognition for STT, SpeechSynthesis for TTS — with a graceful detect-and-fallback when browsers (e.g., Brave on default settings) block Google's speech servers.
  • Infrastructure: AWS EC2 t3.small (ap-south-1), Gunicorn, nginx, Let's Encrypt. Live at debtclear.aryangorde.com.

Challenges we ran into

  • AI tone drift: The financial advisor would say "your situation is challenging" whether the stress score was 20/100 or 80/100. Fixed by mapping the score to four severity tiers (HEALTHY/MODERATE/ELEVATED/SEVERE) and injecting an explicit tone directive that bans certain words for healthy borrowers.
  • Voice on Brave: Brave Shields blocks Google's speech servers, so SpeechRecognition fails with a "network" error. Detect that error code and surface "Voice unavailable — type instead" so the user falls back to text without confusion.
  • Premature settlement detection: The roleplay AI would tag the call as SETTLED while still asking "would you commit to that?" — locking the user out before they could confirm. Fixed by tightening the system prompt to require explicit two-sided confirmation, and by keeping the modal's input live through settled and declined states.
  • Settlement range personalization: Static lookup tables felt fake. Built a separate AI call that takes leverage score + hardship factors and returns a JSON range the script generator uses.
  • WebGL on low-end mobile: At 12% canvas scale the noise shader produced just black. Found 35% as the minimum viable resolution that keeps the nebula visible without melting budget Android GPUs.
  • Amazon Linux Python: AL2023 ships Python 3.9; Django 5 needs 3.10+. Installed Python 3.12 from dnf and rebuilt the venv.

Accomplishments that we're proud of

  • A fully working, live, HTTPS-deployed app — not a mockup. Every feature in this submission works end-to-end at debtclear.aryangorde.com.
  • Five distinct AI capabilities all running on Llama 3.3 70B with deterministic fallbacks: financial advisor, follow-up chat, negotiation script, voice roleplay, settlement letter.
  • The negotiation script is genuinely usable — real creditor language, real dollar amounts, no placeholders.
  • Voice roleplay actually works on Chrome — you can practice the call out loud and an AI creditor pushes back in real time.
  • Zero cost to run during the judging period (free AI tier with key pooling + AWS credits).

What we learned

  • Free-tier LLM APIs are shockingly capable for a hackathon — enough for thousands of full user sessions with proper key pooling.
  • A deterministic fallback isn't a crutch — it's what makes AI-powered products reliable under rate limits.
  • Tone is content. The same advisor with the same numbers reads completely differently when you tell it to "open with confidence" vs. "open with urgency".
  • Empathy beats math when the numbers are bleak. A crisis panel with real nonprofit resources matters more than another chart for someone with a 90/100 stress score.

What's next

  • Plaid integration — pull real balances automatically instead of manual entry.
  • Progress tracker with persistence — log actual payments and watch the simulation update over time.
  • Credit report parsing — upload a PDF, auto-populate all debts.
  • Email reminders & calendar integration — schedule monthly check-ins so the plan doesn't slip.

Built With

Share this project:

Updates