About me

A builder at heart.
With a researcher’s curiosity.

I’m Riddhi, an AI Solutions Architect in Toronto. I design systems that connect language models with business data, APIs, and everyday workflows.

My experience spans enterprise engineering, NLP research, and teaching. At Amgen, I built agents for Legal, Finance, and Pharmacovigilance. Today, I work on AI architecture and automation at Northeast Shared Services.

I care about making technology easier to understand and use. That has shaped my research on efficient, reliable AI—and my work mentoring more than 500 students at Ontario Tech.

My experience ↘My education ↘My projects & papers ↘

How I build AI solutions

Requirements

Define the business problem, user workflow, constraints, and success criteria.

Work experience

Aug 2026 — Present

AI Solutions Architect

Northeast Shared Services

  • Design end-to-end AI solutions that automate business processes and improve operational efficiency.
  • Translate business needs into practical, scalable, and secure technical architectures.
  • Develop AI-powered workflows connecting cloud platforms, enterprise data, APIs, and internal systems.
  • Lead requirements gathering, solution design, research, prototyping, testing, and documentation.
  • Collaborate with business stakeholders, IT teams, subject-matter experts, and technology partners.
  • Evaluate security, governance, data access, traceability, scalability, and long-term support throughout the solution lifecycle.
Nov 2025 — Jul 2026

GenAI Engineer

Amgen

  • Built enterprise GenAI agents using LLMs, RAG, MCP, and agentic workflows.
  • Designed assistants with retrieval, reasoning, tool calling, and workflow automation to improve knowledge access.
  • Developed RAG pipelines for document processing, embeddings, vector search, retrieval optimization, and grounded responses.
  • Integrated enterprise applications, APIs, Teams, Outlook, and internal platforms through MCP-enabled agents.
  • Built applications and services with LangChain, Bedrock, AWS, Node.js, C#, SQL, and React.
  • Improved prompts, orchestration, evaluation, reliability, and hallucination reduction.
  • Partnered with cross-functional teams to deliver secure AI in a regulated pharmaceutical environment.
Enterprise GenAI Agent Solutions ↗
Sep 2023 — Nov 2025

Software & AI Researcher

Software Engineering & Education Research Lab

  • Studied multilingual model realignment and layer sensitivity through ALIGNFREEZE, accepted at NAACL 2025.
  • Awarded funding to optimize LLMs for flaky test detection in resource-constrained environments.
  • Designed FlakyXbert with few-shot Siamese learning; explored fine-tuning, quantization, and distillation.
  • Worked on fairness and data-leakage safeguards for software engineering models.
  • Co-authored HumanEval_T, a modular template-based approach to benchmark generation and code-generation evaluation.
  • Built an agentic research assistant for literature review and knowledge management using local inference and retrieval.
LLM-Powered Agentic Research Assistant ↗FlakyXbert: LLM Optimization for Flaky Test Detection ↗ALIGNFREEZE: Multilingual Model Realignment Analysis ↗HumanEval Data Leakage Mitigation Framework ↗Assessing Data Augmentation-Induced Bias ↗
Jan 2024 — Sep 2024

NLP Researcher

Lee Language Lab

  • Worked on ALIGNFREEZE, analyzing the impact of realignment on multilingual models and internal representations.
  • Contributed to an evaluation framework studying alignment behavior across languages; the resulting work was accepted at NAACL 2025.
ALIGNFREEZE: Multilingual Model Realignment Analysis ↗
Sep 2023 — Apr 2025

Graduate Teaching Assistant

Ontario Tech University

  • Mentored more than 500 students in Software Quality Assurance and Scientific Data Analysis.
  • Developed automated grading systems; the supplied résumé reports a 15% improvement in performance metrics.
  • Led weekly labs and office hours with tailored feedback; the résumé reports 95% student satisfaction.
Jun 2022 — Sep 2022

Globalink Research Intern

Mitacs · Ontario Tech University

  • Worked under Prof. Jeremy Bradbury to develop Threaded Paws, a serious game for learning thread interleaving, data races, starvation, and deadlocks.
  • Designed for second- to fourth-year undergraduate Computer Science and Software Engineering students.
Threaded Paws ↗
Feb 2020 — Jan 2022

Student Technical Fellow

Google

  • Selected as one of 126 participants from more than 15,000 applicants for the Women Engineers / WTEF initiative.
  • Completed a two-year professional development program with live sessions, boot camps, ongoing mentorship, certification, and team projects.
  • Developed problem-solving and computational-thinking skills through the Google-supported TalentSprint fellowship.
Jun 2021 — Aug 2021

MLH Fellow

Major League Hacking

  • Contributed to an open-source contribution platform and developed applications using Ruby on Rails and Jekyll.
  • Collaborated in an agile, cross-functional, distributed team.
  • Worked across the software lifecycle, from conception to deployment.
Aug 2020 — Nov 2020

International Business Consultant

X-Culture

  • Worked in an international virtual team on business solutions for US drone company Dart Drone.
  • Researched markets and developed recommendations for expansion into additional countries.
  • Practiced global virtual collaboration across an internationally distributed team.

Education

MSc, Computer Science

Ontario Tech University

2023 — 2025 · 4.2 / 4.3 GPA

Integrating transformer-based NLP in software testing; automation and efficiency using LLMs.

  • Advanced NLP
  • Research Methodologies
  • Software Design & Project Management

BE, Computer Engineering

University of Mumbai

2019 — 2023 · 9.11 / 10 GPA

Automated Collision Avoidance of Unmanned Maritime Vehicles using Reinforcement Learning. Undergraduate thesis, unpublished.

  • Data Structures & Algorithms
  • DBMS
  • Operating Systems
  • Artificial Intelligence
  • Computer Vision
  • Machine Learning
Thesis project ↗

Projects & publications

Enterprise GenAI Agent Solutions

Connecting AI assistants to the documents, tools, and workflows people already use.

  • RAG
  • MCP
  • LangChain
  • Amazon Bedrock
  • Node.js
  • C#
  • SQL
  • React

Amgen · GenAI Engineer · 2025–2026

Legal, Finance, and Pharmacovigilance teams work across enterprise knowledge and applications, with the access controls and reliability expected in a regulated pharmaceutical environment.

Delivered AI capabilities for enterprise knowledge access and process automation, with security and regulated business requirements included in the solution design.

  • Built agentic assistants with retrieval, reasoning, tool calling, and workflow automation.
  • Developed end-to-end RAG pipelines: document processing, embeddings, vector search, retrieval optimization, and grounded response generation.
  • Built MCP-enabled integrations with enterprise APIs, Microsoft Teams, Outlook, and internal platforms.
  • Developed applications and backend services using LangChain, Amazon Bedrock, AWS, Node.js, C#, SQL, and React.
  • Worked on prompt engineering, orchestration, evaluation, reliability, and hallucination reduction with cross-functional teams.

LLM-Powered Agentic Research Assistant

From a research question to a searchable knowledge base, with local model inference.

  • Python
  • LangGraph
  • Ollama
  • FAISS
  • RAG
  • Docker
  • Tavily

Research tooling · SEER Lab

Literature review involves repeated searching, reading, and organizing. Useful context can get lost between papers and the tools used to read them.

The earlier portfolio reports a 65% reduction in review time in this project. Local inference kept model execution on the host; web search used an external service.

  • Created an end-to-end agent workflow for literature review and knowledge base construction.
  • Used LangGraph to coordinate the workflow, Tavily for web search, and FAISS for retrieval over collected material.
  • Used Ollama for local inference and automated context injection; packaged the tooling with Docker.

FlakyXbert: LLM Optimization for Flaky Test Detection

Finding unreliable tests when training data and compute are limited.

  • Python
  • BERT
  • PyTorch
  • TensorFlow
  • Hugging Face
  • Few-shot learning
  • Siamese networks

SEER Lab · ICST 2025

A flaky test can pass and fail without any code changes. Detecting and classifying these tests is difficult when a team has little historical training data.

Few-shot learning provided competitive accuracy at lower training cost in limited-data settings. The résumé reports 20%+ accuracy improvement and approximately 80% lower GPU usage across the project’s experiments; these are context-specific results.

  • Designed FlakyXbert, a few-shot approach using a Siamese network, and compared it with fine-tuning pretrained language models.
  • Evaluated the methods on FlakyCat and IDoFT to understand the tradeoffs between accuracy, available data, and training cost.
  • The broader résumé project included quantization, distillation, and safeguards for fairness and data leakage in software testing.

An Analysis of LLM Fine-Tuning and Few-Shot Learning for Flaky Test Detection and Classification

Riddhi More & Jeremy S. Bradbury · ICST 2025 · Naples, Italy

Compares fine-tuning with few-shot Siamese learning for flaky test detection and classification. Both approaches can be useful, depending on an organization’s training data and resources.

Read paper ↗ (opens in a new tab)Code & data ↗ (opens in a new tab)

ALIGNFREEZE: Multilingual Model Realignment Analysis

Helping multilingual models learn alignment without losing what they already know.

  • Python
  • PyTorch
  • Transformers
  • Multilingual NLP
  • BERT
  • XLM-R

Multilingual NLP collaboration · NAACL 2025

Realignment can improve transfer across languages, yet reduce performance for languages that differ from the source language used for fine-tuning.

Lower layers were particularly vulnerable to realignment. Freezing them helped preserve performance in languages where full realignment caused degradation.

  • Co-authored a study of layer-wise changes during multilingual model realignment.
  • Compared freezing the lower or upper half of a model’s layers with full realignment.
  • Evaluated three models across 35 languages and four tasks to understand layer sensitivity and language-specific degradation.

AlignFreeze: Navigating the Impact of Realignment on the Layers of Multilingual Models Across Diverse Languages

Steve Bakos, Félix Gaschi, David Guzmán, Riddhi More, Kelly Chutong Li & En-Shiun Annie Lee · NAACL 2025

Introduces selective layer freezing during realignment. Controlled experiments investigate where realignment changes representations and how preserving lower layers can improve multilingual transfer.

Read paper ↗ (opens in a new tab)

HumanEval Data Leakage Mitigation Framework

Fresh code-generation tasks that test problem solving beyond memorization.

  • Python
  • LLM evaluation
  • Combinatorial testing
  • Code generation

SEER Lab · ICST 2025, short papers

Benchmark tasks and solutions may appear in a model’s training data. This can make scores a poor measure of performance on new problems.

Proposed HumanEval_T as an alternative construction of HumanEval, making task variation part of the benchmark rather than relying on a fixed set of questions.

  • Co-authored a benchmark construction method based on reusable task templates.
  • Used combinatorial test design to instantiate new concrete tasks from each template.
  • Balanced variation with comparability: tasks should differ enough to reduce leakage effects and remain similar enough to support fair evaluation.

Addressing Data Leakage in HumanEval Using Combinatorial Test Design

Jeremy S. Bradbury & Riddhi More · ICST 2025 · Short Papers, Vision and Emerging Results

Presents template tasks and combinatorial generation as a way to reduce the impact of contaminated benchmark examples on code-generation evaluation.

Read paper ↗ (opens in a new tab)

Assessing Data Augmentation-Induced Bias

Examining what happens when augmented samples enter both training and testing.

  • Python
  • Machine learning
  • Data augmentation
  • Model evaluation

SEER Lab · FAIRNESS 2025

Augmentation helps when data is scarce or imbalanced, but including synthetic samples in evaluation can change what a model’s score tells us.

The study highlights the need to examine data construction when interpreting model performance, especially when augmented samples enter a test set.

  • Co-authored a case study of augmentation-induced bias in flaky test classification.
  • Examined how augmented training and testing samples affect evaluation, including SMOTE and mutation-based augmentation.
  • Studied how to test for bias instead of assuming augmented data is interchangeable with original observations.

Assessing Data Augmentation-Induced Bias in Training and Testing of Machine Learning Models

Riddhi More & Jeremy S. Bradbury · 1st International Workshop on Fairness in Software Systems · 2025

Uses flaky test classification to examine the effect of augmented samples on model evaluation and to demonstrate a way to assess augmentation-induced bias.

Read paper ↗ (opens in a new tab)Code & data ↗ (opens in a new tab)

Automated Collision Avoidance Using Reinforcement Learning

Reinforcement learning for collision avoidance in a simulated maritime environment.

  • Unity
  • ML-Agents
  • Reinforcement learning
  • Lidar

University of Mumbai · Bachelor’s thesis · 2023

Autonomous navigation needs to account for moving obstacles and wildlife, including potential ship–whale collisions.

Completed “Automated Collision Avoidance of Unmanned Maritime Vehicles using Reinforcement Learning,” an unpublished undergraduate thesis exploring AI-assisted navigation.

  • Built a Unity simulation using lidar-derived sensor inputs to represent the surrounding environment.
  • Used reinforcement learning and ML-Agents to train collision-avoidance behavior.
  • Studied navigation policies in simulated, noisy conditions as part of the undergraduate thesis.

Threaded Paws

A serious game that makes concurrency concepts easier to see and learn.

  • Software engineering
  • Concurrency
  • Game-based learning

Mitacs Globalink · Ontario Tech · 2022

Thread interleaving, data races, starvation, and deadlocks can be difficult to understand through code alone.

Developed Threaded Paws as a teaching-oriented research project during the Mitacs Globalink internship.

  • Worked with Prof. Jeremy Bradbury at Ontario Tech to develop a game for learning concurrency concepts and common pitfalls.
  • Designed the learning experience for second- to fourth-year undergraduate Computer Science and Software Engineering students.

Technical skills

Generative AI & Agents

Machine Learning & NLP

Software Development

Cloud & Delivery

Certifications

  • IBM Machine Learning Specialization · 2022
  • Supervised, Unsupervised & Deep Learning · 2022
  • Applied Scrum for Agile Project Management · 2022