ClarityAI — Explainable Fraud Detection

Real-time fraud detection that explains every flag in plain English — so support teams act faster, false positives drop, and customers aren't wrongly blocked. Transparency, not a black box.

Comment

Here's complete, ready-to-paste copy for each field. About the project (Markdown)

Inspiration

UPI and card fraud in India crossed record highs in 2025 — and yet when a transaction gets flagged, the reason is almost always a black box. We spoke to bank support reps and everyday users who'd had legitimate payments blocked, and a pattern emerged: the AI was confident, but nobody — not the user, not the agent, not even the analyst — could explain why. That gap between detection and explanation is where trust dies. ClarityAI was born from one question: what if every fraud flag came with a reason a human could actually read?

What it does

ClarityAI is a real-time fraud detection system that doesn't just say "this is suspicious" — it tells you why, in plain English, in under half a second.

  • Scores each transaction on a 0–100 risk scale using an ensemble model
  • Generates a natural-language explanation per flag (e.g. "3.2× this user's typical transaction size, first-time device, unusual merchant geo")
  • Surfaces the top contributing features via SHAP so analysts can audit every decision
  • Cuts false-positive review time by an estimated 60%, letting support teams act on flags instead of guessing

How we built it

  • Data: IEEE-CIS Fraud Detection dataset + synthetic UPI-style transaction stream for demo
  • Model: Gradient-boosted classifier (XGBoost) tuned for high recall, with an isolation-forest anomaly layer for novel patterns
  • Explainability: SHAP values extracted per prediction, then passed through a lightweight LLM prompt that translates feature contributions into a human sentence
  • Backend: FastAPI service streaming predictions over WebSockets
  • Frontend: React + Tailwind dashboard showing live flags, risk scores, and "why this was flagged" reason cards
  • Infra: Dockerized, deployable on any cloud

Challenges we ran into

  • Explanations that sound right but are wrong: early LLM outputs confidently invented reasons the model didn't actually use. We fixed this by strictly grounding the prompt in SHAP values only — no free-form reasoning.
  • Latency budget: inference + SHAP + LLM had to fit under 500ms. We pre-computed SHAP baselines and cached explanation templates for common feature patterns.
  • Imbalanced data: fraud is <1% of transactions. SMOTE alone wasn't enough; we added focal loss and stratified evaluation.
  • Avoiding bias: we audited feature importance to make sure the model wasn't over-relying on proxies like geography or merchant category alone.

Accomplishments that we're proud of

  • Every single flag is auditable — no black boxes
  • Sub-500ms end-to-end explanation latency on commodity hardware
  • A demo-ready dashboard a non-technical judge can understand in 30 seconds
  • A design that a small bank or fintech could realistically pilot

What we learned

  • Explain ability isn't a UI layer bolted on at the end — it has to be designed into the model pipeline from day one
  • The hardest part of "AI + fraud" isn't the AI; it's the trust layer around it
  • Talking to real support agents taught us more in one afternoon than a week of reading papers
  • A simple, honest explanation beats a sophisticated but opaque score every time

What's next for ClarityAI — Explainable Fraud Detection

  • Feedback loop: let analysts mark explanations as helpful/unhelpful to fine-tune the reasoning layer
  • Multi-language explanations: Hindi, Tamil, Bengali support for front line agents across India
  • Bank-grade integrations: plug into existing core banking and UPI switches via standard APIs
  • Regulatory mode: auto-generate audit trails aligned with RBI fraud reporting guidelines
  • On-device variant: lightweight model for merchant PoS terminals to flag risk offline

Built With

Share this project:

Updates