Research Engineer @ ZOHO Corp
DevOps Developer ➔ AI Security Engineer ➔ Research Engineer
I'm a mechanical engineer by training, but it turns out the hardest machines are made of linear algebra. Over the past 9 years, I've tackled three distinct, hard problems in AI each one building the foundation for the next:
Pushing the boundaries of how efficiently we can serve massive AI models.
- The Work: Weight quantization (GPTQ, AWQ), KV cache compression, activation-aware methods, and speculative decoding.
- The Impact: Crafting custom CUDA kernels and leveraging vLLM to deliver identical model quality at a fraction of the compute cost.
Securing the perimeter with intelligent, model-driven defenses.
- The Work: Architected ZOHO's core phishing detection engine, evolving from simple URL classifiers to BERT-based analysis and custom multi-head attention models.
- The Impact: Deployed across messaging, mail, EDR, and browser environments. Built robust PII detection for privacy compliance.
Building the plumbing that makes enterprise AI possible.
- The Work: Scaled real-time anomaly detection to handle 10M+ calls/minute and deployed massive 2-30GB production models (ResNet, Seq2Seq).
- The Impact: Handled Redis sharding, Docker provisioning in hostile environments, and major migrations (Spark 1.5.2 ➔ 2.0, Java 7 ➔ 8). Built REST APIs over legacy JAR inference.
|
I write about the concepts and technologies that fascinate me mostly deep dives into AI engineering, optimization, and systems. Check out my latest thoughts in my blog. |
|
"Turned out the hardest machines are made of linear algebra."