Vaibhav Saran.
AI ENGINEER & ML RESEARCHERBASED IN THE UNITED STATES

ENGINEER. RESEARCHER. BUILDER.

Hello, I’m Vaibhav.

Research-minded.
Production driven.

I build AI systems that work beyond the demo.
Agentic workflows, applied machine learning, and
research with a path to the real world.

Explore my work View resume
CURRENTLYAI/ML Engineering Intern at Trek Health
Vaibhav Saran
VAIBHAV SARAN / AI ENGINEER
FROM RESEARCH QUESTIONS TO WORKING SYSTEMSSCROLL TO ENTER

01 / PROFESSIONAL EXPERIENCE

Where I put it into practice.

Building with teams. Solving real problems. Taking responsibility for the system, not just the model.

August 2026 – Present

Trek Health

San Mateo, CA

AI/ML Engineering Intern

Current
  • Cut error-resolution time by 70% for sales teams and customers by deploying a diagnostic skill that traces failed agent runs to priority-ranked root causes, using an evidence-bundle query and mandatory sub-agent drilldown.
  • Improved response time by 87% and reduced model calls by 38% by migrating the platform’s LangGraph orchestration to a REST API architecture, reducing infrastructure costs by 42%.
January 2025 – August 2026

Green AI Labs, University at Buffalo

Buffalo, NY

Research Intern

  • Built an LLM-grounded decision-support framework combining retrieved clinical knowledge with an ensemble offline RL, XGBoost, and TabNet policy over 874K ICU records.
  • Achieved 83% treatment accuracy in research evaluation, a 23 percentage-point gain over the BCQ baseline.
  • Built an admission-time risk-stratification module using HDBSCAN and UMAP over 27,799 patient stays, classifying cases into three validated mortality-risk tiers within minutes to flag high-risk cases for physician review.
January 2023 - June 2024

Innomatics Research Labs

Hyderabad, Telangana, India

Data Scientist

  • Validated a price-aggregation proof of concept for Amazon, BestBuy, and Flipkart, collecting 5,000+ listings through BeautifulSoup and Selenium with automated weekly refreshes and stress-testing under simulated production-scale conditions.
  • Normalized multi-source data into 12 structured attributes and built a distributed PySpark ETL pipeline into PostgreSQL, enabling unified CRUD and cross-brand price analytics.
  • Mentored 150+ learners across seven countries with a 4.7/5 satisfaction rating.

02 / SELECTED ENGINEERING

Ideas, built into systems.

A closer look at the agents, pipelines, and applications I’ve taken from a question to a working product.

01 / MULTI-AGENT RESEARCH2026
Editorial illustration of glass financial research planes, archival records, and a central research lens
RETRIEVE. REASON. RESEARCH.

AGENTIC AI / FINANCIAL RESEARCH

StockPilot AIA research desk, orchestrated.

Built a supervisor-based multi-agent research system combining live stock data, news, and charts through hybrid retrieval, with responses under 10 seconds.

Added a prompt-injection guardrail and an 11.4% data-quality filter. Evaluated retrieval with RAGAS: 0.624 overall, 0.754 faithfulness, and answer relevancy improved from 0.673 to 0.736.

Extended analysis with seven-day Prophet forecasting and archived 3,800+ cold records to AWS S3 with DuckDB fallback. Airflow orchestrates PostgreSQL, MongoDB, and ChromaDB.

<10 secmulti-source research responses
  • LangGraph
  • RAGAS
  • Airflow
  • FastAPI
  • AWS EC2
  • DuckDB
Live projectSource code
02 / AUTONOMOUS RACE INTELLIGENCE2026
Editorial illustration of an unbranded open-wheel car with glass weather, telemetry, and race-report panels
BEFORE THE LIGHTS GO OUT.

AGENTIC AI / MLOPS

F1 Primus AIRace intelligence. Before race day.

Built an autonomous pre-race intelligence pipeline that brings weather, practice telemetry, news, and a Gradient Boosting lap-time model into a single report in under two minutes across 24 rounds.

A LangGraph ReAct agent reasons over FastF1, Open-Meteo, and Tavily sources; MLflow records offline experiments and Prefect handles orchestration and retries.

Containerized the Streamlit application with Docker, deployed it on GCP Cloud Run, and automated deployment with GitHub Actions.

<2 minautonomous pre-race reports
  • LangGraph
  • LangChain
  • FastF1
  • MLflow
  • Prefect
  • Docker
  • GCP Cloud Run
  • GitHub Actions
Live projectSource code

More systems. Different problems.

THE SAME END-TO-END APPROACH
Editorial glass sorting sculpture organizing opportunity cards into a selected shortlist
FROM OPPORTUNITIES TO A SHORTLIST.

WORKFLOW AUTOMATION

Job Search Automation

  • Created a weekday 8 AM or manual n8n workflow that collects LinkedIn jobs through Apify, normalizes fields, and removes duplicate URLs.
  • Claude Sonnet 4.6 scores fit and provides resume suggestions. The pipeline logs 21 fields per job into Google Sheets, routing scores of 70 or above into High Matches while retaining All Jobs.
  • Processes 100 jobs per run, turning job discovery into a repeatable, inspectable pipeline.
  • n8n
  • Claude API
  • Apify
  • Google Sheets
  • Automation
Editorial comparison display of a laptop, phone, tablet, smartwatch, and folding phones on ivory stone plinths
COMPARE. DISCOVER. CHOOSE.

DATA ENGINEERING / RECOMMENDATIONS

Laptrack

  • Built a robust ETL pipeline using PySpark to scrape, preprocess, and load extensive laptop data from Amazon, BestBuy, and Flipkart, ensuring seamless handling of diverse data formats. Dataset published on Kaggle.
  • Developed a feature-rich aggregator application using Flask (backend) and ReactJS (frontend), enabling users to compare laptop buying options across multiple platforms.
  • Implemented a content-based recommendation engine to suggest similar products, enhancing user engagement and providing tailored shopping experiences.
  • Designed and deployed machine learning models to predict laptop prices based on specifications and analyze pricing trends, integrating insights into an admin dashboard for informed decision-making.
  • PySpark
  • Hadoop
  • Flask
  • ReactJS
  • Machine Learning
Watch demoDatasetSource code
Editorial gallery of hand sculptures in distinct fingerspelling poses, framed by sage optical glass
FROM GESTURE TO MEANING.

COMPUTER VISION / ACCESSIBILITY

Hand Sign Detection

  • Built a real-time webcam detector for 10 custom hand signs using TensorFlow Object Detection and SSD MobileNet V2 FPNLite 320×320, connecting visual sign recognition to on-screen text.
  • Created a 300-image dataset by photographing personal examples of each sign and annotating bounding boxes with LabelImg. Used 240 images for training and 60 for testing.
  • Fine-tuned the pretrained detector for 10,000 training steps, then ran predictions on still images and live camera input with bounding boxes, sign labels, and confidence scores.
  • Organized data collection, training, evaluation, and inference in Jupyter notebooks and exported the frozen model. The custom vocabulary includes expressions such as “thank you,” “name,” “what,” “nice,” and “I love you.”
  • Python
  • TensorFlow
  • SSD MobileNet
  • LabelImg
  • Jupyter
Source code

03 / RESEARCH & PUBLICATIONS

Questions worth investigating.

Interpretable clinical decisions and more sustainable machine learning. Work that asks what better could look like.

Editorial clinical glass sculpture with layered records and a decision path, representing interpretable sepsis research
MAKING CLINICAL REASONING VISIBLE.

IEEE ICMLA 2025 / CONFERENCE PAPER

Interpretable sepsis treatment with offline reinforcement learning.

  • Developed an interpretable attention-based offline RL framework with clustering and LLM-generated rationale, evaluated on MIMIC-III and eICU cohorts.
  • The ensemble policy achieved 83% treatment accuracy over 874K ICU records. HDBSCAN and UMAP group 27,799 patient stays into three mortality-risk tiers.
  • Augmented underrepresented fluid and vasopressor examples with synthetic trajectories from diffusion models and a conditional VAE, addressing sparse intervention data during offline training.
  • Awarded third place at CSE Demo Days, University at Buffalo; the work developed into an IEEE ICMLA 2025 conference paper.
83%treatment accuracy in research evaluation
23 percentage points above the BCQ baseline

Kumar P., Saran V., Patel D., Kulkarni N., Vereshchaka A.
Attention-Based Offline RL and Clustering for Interpretable Sepsis. IEEE ICMLA 2025.

  • LangChain
  • Reinforcement Learning
  • HDBSCAN
  • UMAP
  • LLaMA
  • Streamlit
Watch demoSource codeRead paper

With Punit Kumar (lead author), Divyesh Patel, and research collaborators at University at Buffalo.

Editorial glass computing core surrounded by six sustainability panels and an energy ribbon
MEASURE MORE THAN ACCURACY.

NY-BEST 2025 / POSTER PRESENTATION

Green CRISP-ML(Q): a framework for sustainable machine learning.

Extends CRISP-ML(Q) with sustainability checkpoints across all six stages of the ML lifecycle, bringing carbon budgets, resource profiling, and environmental trade-offs alongside model quality.

Proof-of-concept evaluation: Neural Collaborative Filtering on MovieLens-20M, using an Intel Xeon Gold 6126 (48 cores), Quadro RTX 6000 GPU, and 187.4 GB RAM.

93.5%of measured CO₂ came from training
8.57%average GPU utilization during training
0.333kWh total experiment energy
Inside the profiler

A five-phase profiler combines Scalene and CodeCarbon: initialization, execution profiling, parallel monitoring, result aggregation, and output generation. Function decorators and context managers capture CPU, GPU, RAM, energy, and CO₂ measurements, exporting a 43-column dataset with session logs and performance summaries.

  1. Business & data understanding
  2. Data engineering
  3. Model building & training
  4. Model evaluation
  5. Deployment
  6. Monitoring & maintenance

These are observations from one NCF/MovieLens-20M experiment, not measured reductions or a general benchmark. Validation across more AI domains and automated optimization recommendations are future work.

Saran V., Imran A.
Green CRISP-ML(Q): A Framework for Sustainable Machine Learning. Poster, NY-BEST 2025.

  • Sustainable AI
  • Scalene
  • CodeCarbon
  • Environmental profiling
  • NCF
  • MovieLens-20M
View research poster
Preview the original posterOriginal University at Buffalo poster by Vaibhav Saran and Dr. Asif Imran, showing the six-stage framework, five-phase profiler, and NCF experiment findings

04 / THE PERSON BEHIND THE SYSTEMS

An engineer. Always a student.

Vaibhav SaranVAIBHAV SARAN / UNITED STATES

I like the space between
“what if” and “it works.”

I’m an AI engineer with a master’s in Artificial Intelligence from the University at Buffalo. My work moves between agentic systems, data engineering, and applied ML research, including interpretable clinical reinforcement learning and sustainable ML.

I care about what happens around a model: the quality of its data, how it is evaluated, how its decisions can be understood, and what it takes to keep it running.

Outside the core projects, I’ve contributed to open source, mentored learners across seven countries, and continue building a public learning roadmap from Python fundamentals to generative AI.

Explore the learning roadmap
GitHubLinkedInGoogle Scholar

The tools follow the problem.

SKILLS & TECHNOLOGIES

Agentic AI

LangGraph · LangChain · LangSmith · RAG · Multi-Agent Systems · MCP · Claude API · OpenAI API · Gemini · Jev (TypeSafe AI)

Machine learning

PyTorch · scikit-learn · Hugging Face · Offline RL · XGBoost · HDBSCAN · UMAP · VAE · Diffusion Models · NLTK · spaCy · Speech/audio processing

Data & infrastructure

Python · C++ · SQL · PySpark · Airflow · Prefect · PostgreSQL · MongoDB · ChromaDB · FAISS · DuckDB · Hadoop · Snowflake · Databricks · Microsoft Fabric · MySQL · SQLite

Shipping & serving

FastAPI · Docker · GitHub Actions · MLflow · AWS EC2 / S3 · GCP Cloud Run · Nginx · Groq · Cerebras · vLLM · TensorRT · Git · Selenium

Programming languages

Python · C++ · C

Sustainable ML

Scalene · CodeCarbon · Energy & CO₂ profiling · CRISP-ML(Q)

ACADEMIC FOUNDATION

Education

Dec 2025

University at Buffalo

MS in Artificial Intelligence

GPA 3.7 / 4.0
June 2022

JIET Group of Institutions

B.Tech in Computer Science

GPA 3.64 / 4.0

05 / RECOMMENDATIONS

Good work is rarely solo.

A few words from the people I’ve had the opportunity to learn from and build alongside.

NK

Nitin Kulkarni

Ph.D. Candidate, University at Buffalo

Research Mentor - November 2025LinkedIn
He was able to work well in a team, understanding how his work fits into the bigger picture and ensured timely progress to support his team members.
KB

Kanav Bansal

CTO @ Innomatics | ThatAIGuy.com

Direct Manager - April 2026LinkedIn
What differentiates him is his ownership mindset and speed of execution.

06 / START A CONVERSATION

Something interesting
on your mind?

A team to join. A problem to untangle. An idea worth exploring. I’d love to hear about it.

vaibhavsaran8@gmail.com
Schedule a call LinkedIn

Your message is delivered securely via Formspree.