- Designed an AI-ready data architecture integrating Snowflake, Salesforce Health Cloud and Salesforce Marketing Cloud, improving model-ready data consistency by 33%.
- Developed a Journey Progress Monitoring Framework in Snowflake to track patient and prospect movement across journey stages, improving journey completion rate by 42%.
- 01Get the data rightCrawl, clean and structure it.
- 02Build the modelRetrieval-augmented generation and machine learning.
- 03Measure it honestlyIncluding where it does not work.
About
My work sits between machine learning, data analytics and applied AI. On COSMO, a UT Dallas question-answering chatbot, I built the data pipeline and the React frontend and co-built the retrieval layer. In safety-signal, I built a screening tool where code verifies every claim an LLM makes before anyone sees it.
I graduated as a Scholar with High Distinction from The University of Texas at Dallas. I care about the full loop: getting the data right, building the model, and measuring honestly how well it works, including where it doesn't.
Education
2023–2025
M.S. Business Analytics & AI (STEM)
2019–2023
B.E. Computer Science & Engineering
Leadership: Operations Officer, INFORMS student chapter (UT Dallas). General Treasurer, IEEE Student Chapter (Thiagarajar College of Engineering).
Certifications
Oct 2025
Databricks Certified Data Analyst Associate
Oct 2024
Microsoft Certified: Azure AI Fundamentals (AI-900)
Badges and training
Sep 2026
Engineer AI Agents with Agent Development Kit (ADK)
Dec 2023
AWS Academy Graduate: Machine Learning Foundations
Oct 2023
AWS Academy Graduate: Data Engineering
Publications, as co-author
2023
Web Scrapping Tools and Techniques: A Brief Survey
2023
Deep Learning-Based Ocular Disease Recognition
Experience
Where I've worked
- Built a prescription drug market intelligence platform with chat-based analytics via Qlik's native AI API, cutting executive time-to-insight from days of analyst requests to seconds.
- Mapped 180K+ prescribers across 140 specialties in Texas and New York, surfacing top prescribers ($25M–$100M+ in drug costs) and high-spend specialties to inform territory planning and commercial targeting.
- Enabled 3-year brand vs. generic and per-capita market analysis, helping leadership assess generic erosion risk and compare markets on a like-for-like basis.
- Engineered a scalable, low-cost 4-stage Qlik pipeline on free public data, with full lineage and suppression-aware metrics that make every reported figure traceable and defensible.
- Curated a machine learning project repository for a 150+ student course, helping students apply classroom theory to real-world problems; project completion rose 25% over the previous semester.
- Led discussions on decision models and risk analysis for a 63-student cohort in a second course; the grade distribution improved 15% over the previous semester.
- Built a 6-class eye-disease classifier (including glaucoma and cataract) on fundus images using transfer learning on pre-trained CNNs, reaching 83% accuracy; balanced the training classes with SMOTE. Published as co-author in IJETMS (2023).
- Decreased reporting turnaround time by 25% by integrating sales, procurement and warehouse data through ETL workflows in Excel and SQL, improving supply visibility for domestic and international orders.
- Increased on-time delivery performance by 10% by developing dashboards to track KPIs such as order fulfillment, delivery timelines and inventory turnover.
Selected work
Things I've built
Course and personal projects. Code is on GitHub. A dashed outline means the project is still in progress.
safety-signal: adverse-event screening with verified LLM summaries
Personal project, healthcare. A screening tool, not medical advice.
For a drug, it pulls FDA adverse-event reports (openFDA FAERS), flags reactions reported disproportionately often (PRR and chi-square), checks each flagged reaction against the drug's FDA label text, and writes a short plain-language summary with an LLM. The model only sees computed facts, and code verifies its output before anyone sees it: it rejects numbers not in the facts and causal wording, and falls back to a template if needed.
View safety-signal on GitHubMeasured
0.77precision, first held-out run
0.96recall, first held-out run
0.60precision of a call-everything baseline
On 72 human-labelled reaction–drug pairs across 6 drugs (4 held out), label-matching precision was 0.77 and recall 0.96 on the first held-out run, versus 0.60 precision for a call-everything baseline.
In 30 summary runs, 29 passed on the first attempt, 1 after a retry, and none fell back to the template.
Limits
FAERS is voluntary reporting, not incidence data; the evaluation is small with one labeller; the output checks are lexical and structural, not semantic.
COSMO: a RAG chatbot for UT Dallas
Group course project. Not deployed.
Cited answers across 2,000+ pages and 120 UTD subdomains.
A retrieval-augmented chatbot that answers questions about UT Dallas in a few seconds, with a cited source on every factual answer, and refuses out-of-scope questions.
My part: the entire data side (identifying sources, crawling with Scrapy and Playwright, cleaning and summarising pages with a local Llama 3.2 via Ollama, storing them in MongoDB and preparing content for embedding) plus the entire React frontend. I co-built the RAG layer with a teammate who led it.
View COSMO on GitHubLogistics risk analysis: which cities and drivers carry the risk
Group course project (six students), UT Dallas.
Hive tables and Tableau dashboards over 8,000 truck telematics events to rank cities and drivers by risk.
Driver and truck telematics were loaded into a relational schema in Hive (geolocation and trucks tables) and analysed in Tableau to answer five business questions: which cities see events from the riskiest drivers, which cities have the highest rate of events, where drivers travel fastest, the average velocity for each event type, and which drivers have over-speed events.
What it found: Santa Rosa had the most events across all four event types. Occidental had the highest average driver risk factor, 5.33, classed as medium risk. 57 unique drivers had over-speed events, one of them four times.
View the project on GitHubAI Tutor: a multi-agent university department
Personal project. Design complete, build in progress.
Turns any topic into a syllabus, modules and tests, then adapts the next module to the learner's mistakes.
Agents mirror a university department: a Topic Identifier (with a guardrail), a HoD who builds the syllabus, a Counsel who critiques it, a Prof who writes modules and tests, and a TA who verifies them. Sources are discovered automatically from the topic rather than from a hardcoded corpus.
Skills
Pick one to jump to the experience and projects that used it.
All skills
Languages
- Python
- SQL
- R
- Java
- JavaScript
- HTML
- CSS
AI, ML and agents
- Keras
- TensorFlow
- NLP
- LlamaIndex
- RAG
- Vector databases
- LangGraph
- CrewAI
- Prompt engineering
- LoRA / QLoRA fine-tuning
- RAGAS
- Guardrails
- LangSmith
Data and cloud
- AWS
- Microsoft Azure
- IBM Cloud
- Snowflake
- MongoDB
- Hive
- Hadoop
- Spark
- Salesforce
- vLLM
Visualization and analytics
- Tableau
- Power BI
- IBM Cognos
- Qlik
- Excel
Developer tools
- Cursor
- Claude Code
Contact
Say hello.
If you have a question or an opportunity, email me and I'll reply.
snandhakumarraj@gmail.com
Based in Dallas, TX, and open to relocation.