ReBank

Reimagining banking, not your traditional way around, but instead a way that enables automation and smooth UI with personalized AI agents.

Comment

Regional banks told us their customers still call about the same three things “Why is my balance off?”, “Where did this card charge come from?”, and “Can you flag this fraud?”—yet agents sift through five different dashboards and stale data dumps to answer. We built ReBank to collapse those silos into one trustworthy, voice-driven surface that can actually keep up with a customer’s finances in real time.

## What it does

  • Spins up four AI specialists—Nebula for spending, Atlas for long-term planning, Sentinel for security, and Nova as an all-purpose concierge—and deterministically routes every request to the right one in under 100 ms.
  • Pre-computes and caches the financial context (balances, category spend, anomaly scans, income trends) with an 18× speed-up by running 11 Nessie- powered tools in parallel and layering memory + file caches.
  • Returns a conversational response (and optional ElevenLabs voice) through a React Native interface that also shows dashboards, analytics, tagging and alerts sourced from the same context.
  • Keeps working even if upstream APIs hiccup thanks to a one-hour persistent cache and graceful fallbacks.

## How we built it

  • FastAPI backend with asyncio orchestration, deterministic keyword routing (194 unique triggers validated by scripts), and multi-layer caching (memory + JSON buffer) feeding the agents.
  • Tooling around the Capital One Nessie sandbox to hydrate transactional data, run anomaly detection, calculate savings rates, and aggregate everything for the agents.
  • A React Native (Expo) mobile client that lets users pick an agent, chat, and explore dashboards/analytics in a responsive layout.
  • ElevenLabs ConvAI integration points that package the cached context so each agent can answer in-character without re-querying the bank.

## Challenges we ran into

  • Designing a zero-ambiguity routing matrix so “save more” goes to the budget coach while “save for retirement” hits the investment advisor; we burned time building the keyword validator to catch collisions.
  • Managing response latency—sequential Nessie calls were 1.5 s+, so we had to re-architect the tool executor around asyncio.gather, timeouts, and error fan-in.
  • Keeping the cache trustworthy: we needed a fallback strategy that serves stale-but-safe answers instead of failing open when upstream services blip.
  • Faking real accounts on the Nessie sandbox; we scripted data hydration to make demos feel lived-in without leaking real PII.

## Accomplishments that we're proud of

  • A fully deterministic multi-agent router with explainable reasoning and zero keyword overlap across 194 triggers.
  • 95 % cache hit rate with responses dropping from ~1.5 s to 50–80 ms, all while keeping fraud checks and spending analyses current.
  • A mobile experience that mirrors the voice story—agent chat, analytics, tagging—so users never leave the ReBank ecosystem.
  • Plug-and-play ElevenLabs layer: once agent IDs are in the .env, every response upgrades from text fallback to voiced personalities.

## What we learned

  • Deterministic orchestration beats fuzzy AI routing when your users need compliance-grade audit trails.
  • Pre-compute everything you can; giving the model context snapshots is faster and safer than letting it hit core banking systems live.
  • Banking data storytelling matters—clear dashboards and category insights help users trust the assistant’s decisions.
  • Agent personality and guardrails must be encoded in code and configs, not just prompts, if you want consistent behavior.

## What’s next for ReBank

  • Wire up real ElevenLabs agents for live voice conversations and push-to-talk inside the app.
  • Move the routing keywords into an operations UI so compliance teams can approve changes without redeploying.
  • Layer in proactive alerts (e.g., “spending spike in dining” or “new unusual transaction”) using the existing anomaly tooling.
  • Expand data sources beyond Nessie—plaid-style aggregators, bill pay, credit scoring—to deepen the context each agent can reason over.

Built With

Share this project:

Updates