Tejas Goyal

Returns fraud costs D2C brands 5–10% of revenue. Most Shopify merchants have zero tooling to detect or prevent it.

Built end-to-end: fraud scoring engine, Shopify integration, merchant dashboard, rule configuration. Sole technical owner.

• 100+ B2B users on production
• Processing real transaction volume across multiple Shopify stores
• Custom rule engine with configurable fraud signals

Onboarding UX matters more than model accuracy at early scale. Merchants churned when setup took >10 minutes, regardless of detection quality.

Most AI tooling is built for demos, not production. Real products need sub-200ms responses and actual uptime guarantees.

Early engineering hire. Worked across the stack on latency-sensitive pipelines, deployment infrastructure, and system reliability.

• Production systems serving real users with strict latency and uptime constraints
• Real-time voice and inference pipelines
• Full-stack ownership: infra, backend, integrations

Notification infrastructure is deceptively complex. Ordering, deduplication, retry logic, multi-channel fan-out all break at scale.

Designed and built the queue architecture, retry/backoff logic, and channel routing layer.

• BullMQ + Redis for reliable async job processing
• Multi-channel delivery (email, push, in-app) with per-channel retry policies
• Dead letter queues and observability built in from day one

Queue depth monitoring is non-negotiable. Without it, you discover backpressure problems only when users complain.

Manufacturing execution systems in aluminium wire rod plants lack real-time quality feedback. Operators catch defects too late, and parameter drift goes undetected across production lines.

Led the technical architecture and delivery under 36-hour hackathon constraints. Built the ML inference pipeline, parameter monitoring layer, and operator-facing dashboard.

• Won Smart India Hackathon 2024 (national-level)
• Deployed on the manufacturer's private intranet, running in production
• Real-time parameter monitoring with ML-backed anomaly detection across manufacturing units

Industrial deployment is a different game. The hardest part wasn't the model, it was fitting into existing plant infrastructure and operator workflows.