Skip to content

ROADMAP.md

Latest commit

 

History

History
57 lines (41 loc) · 2.46 KB

ROADMAP.md

File metadata and controls

57 lines (41 loc) · 2.46 KB
Raw

RailHawk Roadmap

Public roadmap. Dates are targets, not guarantees.

v0.6.0 — Current release

  • Multi-rail transaction schema (UPI / AEPS / DMT / Cards / Loans)
  • 3-layer scoring stack: Rules → 6 parallel ML models → Meta-learner
  • 466-feature Redis pipeline (Tier 1 velocity + Tier 2 profiles)
  • FastAPI inference server (P95 < 100ms target)
  • Specialist models: Agent Risk Model, DMT Mule Detector, GNN Ring Detector
  • Cross-rail risk propagation
  • 7-phase training pipeline with Optuna HPO and MLflow tracking
  • Drift detector (PSI)
  • PCI-DSS + Aadhaar Act §29 + PMLA compliance guardrails
  • GitHub Actions: CI (lint + security + tests), Docker → GHCR, Release automation
  • OSS community files: CONTRIBUTING, SECURITY, CoC, issue templates, Dependabot

v0.7.0 — Next

  • ONNX warmup at startup — dummy inference on each ONNX model during lifespan() to eliminate first-request JIT cost (~200ms)
  • Champion/challenger routing — wire hash-split routing in api/routes/transaction.py; shadow scoring already works, traffic split needs wiring
  • Drift → auto-retrain trigger — connect drift_detector.py PSI threshold breach to RETRAIN_FLAG_PATH watcher
  • Helm chart — charts/railhawk/ full Deployment + HPA + PVC + ConfigMap + Secret manifests
  • Self-hosted GPU runner — register runner on RTX 4000 host; activate train.yml workflow
  • Pre-trained weights on HuggingFace — first public model release (synthetic-data weights)

v0.8.0 — Future

  • Kafka chargeback feedback loop — activate kafka/chargeback_consumer.py; wire confirmed fraud labels back to retraining
  • Multi-tenancy — per-client_id thresholds, model routing, and audit isolation
  • SBOM — Syft SBOM generation in build.yml, attached to GitHub Release
  • OpenSSF Scorecard — badge in README, automated weekly scoring
  • Codecov integration — coverage trend tracking
  • GitHub Pages docs site — auto-publish docs/ via MkDocs + Material theme

Backlog (unscheduled)

  • Real-data fine-tuning guide (bring-your-own labeled dataset)
  • Kubernetes operator for model lifecycle management
  • Grafana dashboard bundle (pre-built JSON for fraud monitoring)
  • Load test benchmark results in docs
  • Multi-region deployment guide (active-passive for data localisation)

Contributing

Want to work on any of these? Open a feature request issue or comment on an existing one. PRs welcome — see CONTRIBUTING.md.