About
Hello, I'm Kaxit Pandya — a GenAI and Machine Learning Engineer based in
Ottawa, Canada, with 3+ years in AI/ML and 2 years building and shipping production LLM systems. I hold a
Master of Engineering in Computer Engineering (AI specialization) from the University of
Waterloo.
I currently work as a GenAI Engineer at TrendAI (Trend Micro), where I fine-tune
self-hosted large language models for AI red-teaming. My SFT → DPO/RPO → GRPO pipeline took a
Llama-3.1-8B attacker model past a Claude Sonnet 4.5 baseline at zero inference cost, and a 4B Gemma-3
security judge I fine-tuned cut missed attacks by 91%.
Before that, I architected an enterprise multimodal GenAI platform for the Bank of Montreal
(BMO) on AWS Bedrock — a LangGraph multi-agent orchestrator routing across 15 specialized agents
for three business units — and built GraphRAG, agentic, and OCR pipelines at Defai Digital and Genellipse.
My research spans the Indian Space Research Organisation (ISRO), IIT
Gandhinagar, National Chung Cheng University in Taiwan, and the University of
Waterloo — covering radar pulse classification, GAN-based inverse design of 2D materials, and federated
learning for water quality prediction, with four papers published or under review.
- Location: Ottawa, Ontario, Canada
- Phone: +1 365 998 6883
- Email: kaxitpandya@gmail.com
- Focus: GenAI · LLMs · Agents
Interests
Generative AI
LLM Fine-Tuning & RLHF
Multi-Agent Systems
MLOps & Cloud AI
Education
Master of Engineering in Computer Engineering (AI Specialization)
University of Waterloo · Waterloo, Canada
- Generative AI
- Large Language Models (LLMs)
- Machine Learning
- Deep Learning
Bachelor of Technology in Computer Engineering
Pandit Deendayal Energy University · Gandhinagar, India
- Artificial Intelligence
- Machine Learning
- Data Structures & Algorithms
- Database Management Systems
Publications
Federated Ensemble based Water Quality Classification on Drone Images (Submitted)
An Intelligent Fashion Object Classification Using CNN — EAI Endorsed Transactions on Internet of Things (Published)
Water Quality Prediction using Machine Learning — Springer (Published)
Advancement in Phishing Attacks and Machine Learning Approaches (Submitted)
Certifications
– AWS Certified Machine Learning Engineer – Associate, Amazon Web Services
– LangChain Mastery and OpenAI Python API Bootcamp, Udemy
– Generative AI with LLMs, Machine Learning Specialization, and Deep Learning Specialization, DeepLearning.AI
– Data Analyst with Python, IIT Madras via NPTEL
– The Web Developer Bootcamp 2023, Udemy
Experience
TrendAI (Trend Micro)
GenAI Engineer · Ottawa, Canada
– Present
- Fine-tuned a self-hosted Llama-3.1-8B attacker model (LoRA r=128, 335.5M trainable parameters) on a 334K-example adversarial corpus through an SFT → DPO/RPO → GRPO pipeline on a single H100, shipped to production in Trend Micro Vision One at $0/call as a replacement for frontier-model red-teaming
- Achieved a 30.8% attack success rate versus 23.5% for the Claude Sonnet 4.5 baseline (z = 2.04) on a fixed 5-seed × 120-attack benchmark across six vulnerability objectives, with ~8× lower variance across seeds (std dev 1.0 vs 8.0) and 0% refusals versus 50–72% for aligned baselines
- Lifted attack success rate from 19.3% to 30.8% (+60% relative) through RL post-training — diagnosed plain DPO stuck at chance, fixed it with RPO, then ran GRPO against a live judge reward — cracking two dead-zone objectives (Sensitive-Data 0% → 27%, Hallucinated-Entities 7% → 32%); built a 5-seed evaluation harness with pairwise z-tests and a deterministic canary cross-check agreeing within 0.3 pp of the LLM judge
- Fine-tuned a 4B-parameter Gemma-3 security judge (QLoRA) on 11.8K adversarial examples, cutting missed attacks by 91% (58 → 5 across 643 held-out cases) and raising F1 from 0.79 to 0.95 over the production judge; lifted package-hallucination detection 7× (F1 0.13 → 0.89) and reasoning-trace secret detection from 41% to 100%
Bank of Montreal (BMO)
GenAI Developer · Toronto, Canada
–
- Architected an enterprise multimodal GenAI platform on AWS Bedrock with a LangGraph multi-agent orchestrator routing across 15 specialized agents — Pydantic structured outputs, guardrails, and real-time moderation — serving three internal business units
- Built an Agentic SDLC Orchestrator spanning Business Analyst (BRD), Developer (solution plan, code generation), and QA (test plans, Selenium automation) agents with knowledge-base grounding and human-in-the-loop approvals
- Designed and deployed cloud-native GenAI microservices and MLOps-ready workflows using AWS CDK, serverless Lambda, and boto3/Bedrock APIs, enabling concurrent real-time inference at scale; built an in-house Figma plugin for automated Figma-to-React Native/Angular visual-to-code generation
Defai Digital
AI Engineer · Markham, Ontario
–
- Engineered a GraphRAG pipeline integrating Memgraph with LangChain, enabling structured knowledge-graph retrieval for enhanced multi-hop reasoning and contextual search; executed document extraction using MinerU and fine-tuned YOLO for precise segmentation
- Deployed the pipeline as a microservice, incorporated Cache-Augmented Generation with GPTCache and Redis, and orchestrated a chain-of-agents using agent frameworks to automate complex, multi-step AI workflows
Genellipse Inc.
AI/ML Data Engineer · Mississauga, Ontario
–
- Built an OCR pipeline for processing PDF and image files using Google Cloud Vision to extract data from U.S. death certificates at scale
- Used BERT-based transformer models to map key-value pairs across non-standardized certificate layouts from all 50 U.S. states, raising field-extraction accuracy to 96% over the rule-based baseline, and implemented LLMs to interpret extracted information for client reporting and decision-making
University of Waterloo — Dr. David Hammond
Research Assistant · Waterloo, Canada
- Implemented an automated web-scraping system using Selenium to extract detailed grocery-item information from Walmart, Sobeys, and Target, with integrated export to Excel for analysis
- Built an image download module to systematically retrieve product images from scraped sites and applied OCR techniques to extract textual information from them
Indian Institute of Technology (IIT) Gandhinagar
Research Intern · Gandhinagar, India
–
- Designed a deep learning model using Generative Adversarial Networks (GANs) for the inverse design of 2D heterostructure materials, enabling efficient exploration of material properties
- Conducted post-processing of atomistic simulation data in Python to extract and visualize material properties
- Converted QuantumATK functionality into Python scripts to automate the creation of heterostructure configurations
National Chung Cheng University — TEEP@AsiaPlus
Research Intern · Chiayi, Taiwan
–
- Implemented a Federated Learning model to predict water quality parameters from drone-captured images, preserving privacy across geographically distributed datasets, and authored a research paper detailing the methodology and results
- Executed robust aggregation strategies including FedAvg, FedAvgM, FedMedian, and FedProx using the Flower framework, handling non-IID data distributions and system heterogeneity
- Applied Explainable AI (XAI) techniques such as SHAP to interpret predictions, highlighting parameter significance and improving transparency
Space Applications Centre, Indian Space Research Organisation (ISRO)
Research Intern · Ahmedabad, India
–
- Developed deep neural networks to classify Pulse Radar Interval Modulation
- Processed noisy, raw radar pulse data captured directly from ISRO satellites
- Achieved 94% classification accuracy on complex Pulse Repetition Interval (PRI) patterns — staggered, jittered, dwell-and-switch, sliding, linear, and sinusoidal
HumanAnsys
Software Developer Intern
–
- Worked with cloud systems, architecture, and core AWS services
- Deployed a video-calling website using AWS EC2, S3, and Lambda
Hybrid Utopia
Web Developer Intern
–
- Built websites from scratch using modern frameworks including Node.js and Express
- Implemented Create, Read, Update, Delete (CRUD) operations with MongoDB
Projects
Fine-Tuning LLMs on Fleet Maintenance Data
Fine-tuned large language models for predictive fleet maintenance using PEFT techniques such as LoRA.
DeepFake Audio Detection System
Preprocessed audio with Librosa and classified clips as deepfake or human-generated using pre-trained models.
Above Ground Biomass Prediction
Hyperparameter-optimized sparse autoencoder embeddings fed into tuned neural network and random forest regressors.
Skills
Programming Languages
Generative AI & LLMs
Machine Learning & Deep Learning
Cloud & MLOps
Databases & Vector Stores
Data, CI/CD & Tools
Extra-Curricular Activities
Technical Project Manager — Waterloo AI Institute (WAT.AI), University of Waterloo
–
- Built deployable AI models for onboard satellites, focusing on semantic segmentation of methane plumes using hyperspectral ML models
- Led AI research projects, managing teams and ensuring timely, high-quality deliverables in collaboration with industry partners
Engineering Representative — Graduate Student Endowment Fund (GSEF), University of Waterloo
–
- Reviewed and evaluated funding applications, ensuring strategic allocation of university resources across academic disciplines to graduate student projects and initiatives
- Provided strategic input on funding decisions to enhance the academic and professional development of graduate students
Chairperson — GeeksforGeeks Student Chapter, PDEU
–
- Managed events and workshops at regular intervals
- Led a team of 50 members and developed the coding culture at the university
Web Development Head — SnT, Science & Technical Club of PDEU
–
- Created the website for the university's flagship event, Tesseract
AWS DeepRacer Contributor — Amazon Web Services
- Placed in the top 10% of 4,000 students in the AWS DeepRacer Student League
- Wrote a reward function for training a DeepRacer car with a reinforcement learning model on AWS
Open Source Contributor — Hacktoberfest 2022
- Made 6+ open source contributions on GitHub using HTML, CSS, JavaScript, Tailwind CSS, Node.js, and MongoDB
Contact
© 2026 Kaxit Pandya · Ottawa, Canada · View résumé