Experience

  1. – Present

    Associate Consultant, AI Engineering

    Applexus Technologies, Seattle, WA

    AI engineer on client-facing Databricks engagements. I scope the use case with the business team, build the agents, deploy them in the client's Azure Databricks workspace and stay through go-live.

    • Design and ship domain-specific AI agents for finance, spend analytics, supply chain and predictive maintenance, giving business teams natural-language access to governed data through Unity Catalog.
    • Build and maintain the multi-agent orchestration in LangGraph: a supervisor over domain specialists, planner and executor on different model tiers to control cost, deterministic routing where no model is needed, and human review before write-back.
    • Deploy on Databricks Apps and Model Serving with FastAPI back ends and React or Dash front ends, with MLflow tracing on every run.
    • Build EULR, Applexus's data and use-case discovery accelerator for Databricks: a LangGraph agent layer over a FastAPI back end and a React front end.
    • Run Databricks self-service training sessions for client data teams.

    Python, LangGraph, LangChain, MLflow, Azure Databricks (Unity Catalog, Delta Lake, Databricks Apps, Model Serving), PySpark, FastAPI, React

  2. –

    Trainee Consultant, Data Science & AI

    Applexus Technologies, Seattle, WA

    Joined after my M.S. and was promoted to Associate Consultant after nine months.

    • Built production agents on Azure Databricks with LangGraph and MLflow for finance, spend analytics, supply chain and promotions, and fixed the state-management and over-planning problems that surfaced before deployment.
    • Engineered an internal AI-powered EDA platform in Python, Node.js and Electron with 50+ statistical tests, local LLM interpretation with Ollama and automated PPTX and PDF reports, cutting EDA cycle time by about 50%.
    • Ran AI and ML feasibility assessments for 10+ enterprise use cases, including Databricks and Snowflake platform comparisons, and built spend-analytics and demand-forecasting pipelines on Azure Databricks.
    • Represented Applexus at an on-site client hackathon and at Databricks Data + AI Summit 2026.

    Python, Azure Databricks, LangGraph, MLflow, Ollama, Node.js, Electron

  3. –

    Data Analyst

    EcoServants Project, Remote, USA

    Built a sponsorship data-analysis framework that turned partnership data into targeting insights, increasing new-partner identification by 40% and donor-targeting accuracy by 35%.

  4. –

    Data Scientist, Capstone Intern

    Honda Research Institute USA, Remote

    My M.S. capstone with Honda Research Institute's 99P Labs, as team lead.

    • Built and deployed a GPT-4-based data preprocessing pipeline for transformation, imputation and anomaly detection, with MLOps practices, reducing manual effort by about 60%.
    • Benchmarked GPT-4, Mistral 7B, Falcon 7B and LLaMA 3 on structured-output tasks and chose the model per task, improving processing accuracy by about 30% over the baseline.
    • Built a Streamlit front end with Groq-hosted Mixtral-8x7b for interactive exploration of the processed data.

    Python, GPT-4, LLaMA, Mistral, Falcon, Groq, Streamlit

  5. –

    Data Scientist Intern

    Pucho Digital Health Inc., Remote, India

    Modified a ResNet50 architecture for audio-based COVID-19 detection and shipped it to production, with 4% better precision and 15% better detection accuracy. Added MFCC spectral-subtraction noise reduction, about 20% better on noisy inputs.

  6. –

    Data Science Research Intern

    AI & Robotics Center, VIT-AP University, India

    Deep learning and computer vision research that led to a published patent application, an IEEE paper and a PVSEC-31 talk. See research.

Selected work

Production AI agents from my work at Applexus, an open-source tool for coding agents, and two data science projects with their results.

Applexus Technologies, client work

Governed AI agents on Azure Databricks

Domain agents for finance, spend analytics, supply chain and predictive maintenance that answer business questions in natural language over governed data in Unity Catalog. I scope each use case with the business team, build the agents and deploy them in the client's workspace through go-live.

A LangGraph supervisor routes each question; planner and executor run on different model tiers to control cost, and a person reviews anything before it is written back.

Client work, so there is no public code. The diagram shows the architecture.

LangGraph, LangChain, MLflow, Unity Catalog, Databricks Apps, Model Serving, FastAPI, React

Architecture of the multi-agent system A business question goes to a supervisor agent. The supervisor routes deterministically when no model is needed, or hands the question to domain agents for finance, spend analytics, supply chain and predictive maintenance, with planner and executor on different model tiers. The agents read governed data through Unity Catalog. A person reviews anything before it is written back, and MLflow traces every run on Databricks Apps and Model Serving. Business question Supervisor agent Finance Spend analytics Supply chain Predictive maintenance Unity Catalog: governed data Human review Write-back routes each question Deterministic route, no model Planner and executor tiers MLflow tracing on every run Databricks Apps, Model Serving

CU Boulder DaSSA Hackathon, top-5 finish, 2023

Predictive maintenance for aluminum recycling

A Random Forest flags bad decoating cycles in aluminum scrap recycling from furnace sensor readings, so maintenance can step in before costly dross builds up. Tuned for recall, because a missed bad cycle costs far more than a false alarm.

Caught 469 of 498 bad cycles on held-out data (94% recall) at 75% precision, with 0.988 AUC.

Python, scikit-learn, XGBoost, SMOTE, GridSearchCV

Case study Code
Confusion matrix on 3,239 held-out cycles: 2,588 good predicted good, 153 false alarms, 29 bad cycles missed and 469 caught.

Text-mining project, 2025

Social media misinformation, an NLP study

1,054 news articles on whether platforms should be held responsible for misinformation, with stance labels from a BERT model, LDA topics, clustering and classifiers that predict an article's stance.

SVM reached 68.7% on three-way sentiment and 62.6% on four stances, against a 50% majority-class baseline.

The labels come from a model, not human annotators, so accuracy measures agreement with those labels.

Python, scikit-learn, NLTK, Hugging Face Transformers, R

Case study Code
Accuracy on held-out articles: SVM RBF on sentiment 68.72%, SVM RBF on four stances 62.56%, linear SVM 59.24%, Naive Bayes 56.40%, best decision tree 51.66%.

More projects

Research

About

I studied computer science with a focus on data analytics at Vellore Institute of Technology, where research at the AI & Robotics Center led to a published patent application, an IEEE paper and a PVSEC-31 talk. After a data science internship at Pucho Digital Health, I earned an M.S. in Data Science at CU Boulder and led my capstone team for Honda Research Institute. I joined Applexus Technologies in November 2025 and was promoted to Associate Consultant in AI Engineering in August 2026. I'm based in Redmond, WA.

Education

  • University of Colorado Boulder

    M.S. Data Science

    GPA 4.0 / 4.0, Aug 2023 – May 2025

  • Vellore Institute of Technology

    B.Tech Computer Science (Data Analytics)

    GPA 9.07 / 10, Jul 2019 – May 2023

Skills

Agentic AI and LLMs
LangGraph, LangChain, multi-agent orchestration, MCP, RAG, Databricks Vector Search, prompt engineering, LLM evaluation, Ollama, LiteLLM, Hugging Face
ML and MLOps
PyTorch, TensorFlow, scikit-learn, XGBoost, Keras, NLP, computer vision (OpenCV), MLflow (tracking, registry, tracing), Databricks Mosaic AI, FastAPI, Docker, CI/CD
Data and cloud
Azure Databricks, Unity Catalog, Delta Lake, PySpark, SQL, Snowflake, AWS (S3, EC2, SageMaker), ETL pipelines
Languages, tools and BI
Python, SQL, R, Java, JavaScript, MATLAB, Bash, Git, Power BI, Tableau