// Staff Machine Learning Engineer @ Zscaler

Hi, I'm Vishanth.

5+ years in ML & software
M.S. UCLA
B.Tech NIT Trichy

I build & fine-tune AI guardrails against jailbreaks.

Staff Machine Learning Engineer at Zscaler, leading ML R&D for AI Guard. I build and fine-tune guardrail models that stop jailbreaks and prompt injection, and I optimise GPU inference for 2× throughput at half the latency. Previously Cornami, Disney and Citi.

Get in touch Résumé See my work

🤗

Now: AI Guard ML R&D

📍 San Jose, California

ex-Citi · Disney · Cornami

~$2M Saved per year through inference optimisation

2× Requests per second after GPU optimisation

50% Lower inference latency

70B+ Parameter LLMs optimised on DGX H100

01 · About

Making LLMs safe and fast.

I'm a machine learning engineer working where model research meets production systems. At Zscaler I build and fine-tune the AI guardrail models in AI Guard, which stop jailbreak and prompt-injection attacks before they reach an LLM. I also make those models fast and cheap at scale: my GPU inference optimisations doubled throughput, halved latency and save roughly $2M a year.

Before that I built and benchmarked inference engines for frontier open models (Llama 4, DeepSeek V3/R1) on DGX H100 at Cornami. I ran causal-inference research on payments at Disney, built RAG products at traversaal.ai, and spent two years building AIOps systems at Citi.

I hold an MS in Business Analytics (Data Science) from UCLA Anderson and a B.Tech from NIT Trichy, so I care as much about the business impact as the model metrics.

🛡️

AI Guardrails & Security

I build and fine-tune AI guardrail models that detect jailbreaks, prompt injection and data leakage in Zscaler's AI Guard.

⚡

Inference Optimisation

GPU optimisation with TensorRT, ONNX and vLLM: 2× RPS, ~50% lower latency and ~$2M/year saved.

🧩

Distributed Serving

Tensor & data parallelism, FSDP, KV caching, flash attention and quantization for 70B+ models.

🔁

MLOps

Taking models from research notebooks to reliable, monitored production systems.

02 · Experience

Where I've worked

Click a role to expand the details.

Zscaler

San Jose, CA
Staff Machine Learning Engineer Oct 2026 – Present
- Leading the AI Guard ML R&D initiative, Zscaler’s product that protects LLM applications
- Build and fine-tune AI guardrail models that detect jailbreak and prompt-injection attacks (encoder/decoder fine-tuning, LoRA)
- GPU inference optimisation (TensorRT, ONNX) that delivered 2× requests per second and ~50% lower latency, saving ~$2M per year
- Benchmarking GPUs, including AMD, for self-hosting and optimising ML models
Senior Machine Learning Engineer Dec 2025 – Oct 2026
- Researched and developed state-of-the-art detectors for prompt injection, jailbreaks and LLM data leakage by fine-tuning encoder and decoder models
- Scaled and optimised model inference with ONNX, TensorRT and vLLM on GPUs/TPUs
- Built MLOps for the AI Guard detectors, taking them from research to production
- LLM Security
- Fine-tuning
- TensorRT
- ONNX
- vLLM
- MLOps
Cornami, Inc.

San Jose, CA
AI/ML Software Engineer Feb 2025 – Dec 2025
- Developed and validated LLM engines for Llama 4, DeepSeek V3 / R1 and Llama 3.3 70B, optimising inference with KV cache, flash attention, quantization, tensor and data parallelism
- Benchmarked models across TensorRT-LLM, vLLM, llama.cpp, Hugging Face and Ollama, turning performance data into engineering and business recommendations
- Researched inference strategies for in-house Fractal cores and ran 70B+ models on DGX H100 with distributed techniques (TP, FSDP)
- LLM Inference
- DGX H100
- TensorRT-LLM
- vLLM
- FSDP
The Walt Disney Company

Los Angeles, CA
Data Scientist, ML Research (Co-op) Sep 2024 – Dec 2024
- Built a causal inference model measuring the impact of each payment provider by region, optimising payment processing and processor fees, reducing costs by 10%
- Engineered ML models for churn, interchange fees, payment failures and chargebacks
- Executive KPI dashboard over a 10-billion-row payments pipeline; optimised Snowflake data models and ETL
- Causal Inference
- Snowflake
- Experimentation
traversaal.ai

Los Angeles, CA
Machine Learning Engineer Jun 2024 – Aug 2024
- Built an enterprise multimodal RAG chatbot (images and slides via OCR) with semantic search for traversaal-pro
- Designed the end-to-end ingestion pipeline (upload → S3 → embeddings → Qdrant) with retrieval, ranking and auth
- Fine-tuned Llama models for ranking and retrieval; semantic chunking gave 20% better retrieval accuracy
- Shipped FastAPI services on AWS as API-as-a-Service for clients
- RAG
- LangChain
- FastAPI
- Qdrant
- AWS
UCLA

Los Angeles, CA
Graduate Teaching & Research Assistant Jan 2024 – Nov 2024
- Built open-source statistics libraries for bioscience (power analysis, hypothesis testing, resampling), published as resamp on PyPI
- Teaching assistant for LS 40 Statistics in Bioscience and MGMT 264A Market Research; grader for NLP in Social Science
- Open Source
- Statistics
- Teaching
Citi

Chennai, India
Software Engineer 2, AI/ML Innovation Jul 2022 – Sep 2023
- Tech Innovation Manager for the APAC–EMEA Business Command Center and HNI team at Citi Global Wealth
- Built ML-driven bots and dashboards (NLP, regression, text mining) to predict and monitor major incidents, saving ~$250K per year
- Centralised monitoring that reduced downtime by 10%; sentiment-analysis dashboard for the Citibank app
Software Engineer Jul 2021 – Jul 2022
- Tech lead at the AIOps Innovation Lab, where I designed the stack (Angular, Node.js, Flask) and Jenkins CI/CD, and set up production infrastructure
- Integrated Splunk, ServiceNow and IBM Cognos via REST into a central pipeline for incident forecasting dashboards
- Introduced Agile/Scrum SDLC, saving 100+ person-hours
🏆 Gold Gratitude Award · Bronze & Copper Gratitude Awards · Client Obsession Award · Bravo Award
- AIOps
- NLP
- Full-stack
- CI/CD
VEvolve Software

Tamil Nadu, India
Machine Learning Engineer (Intern) May 2020 – Aug 2020
- ML models (text mining, web scraping, random forests) to assess PHP application security vulnerabilities, generating a $30K revenue share
- Cybersecurity
- ML
Concordia University, John Molson School of Business

Montréal, Canada (remote)
MITACS Globalink Research Intern Apr 2020 – Jul 2020
- Built a heuristic assortment-optimisation model with Prof. Satyaveer, increasing modelled profit by 6% for Walmart Canada
- Optimisation
- Research
IIT Madras

Chennai, India
Summer Research Intern, Data Analysis May 2019 – Jul 2019
- MATLAB data-analysis model for a single-cylinder Gasoline Direct Injection engine at the NCCRD lab, detecting abnormalities in engine parameters
- Research
- MATLAB