01◆ F1 Prediction Engine
Data ·Python / XGBoost / Monte Carlo
F1 Prediction Engine
A triple-model ensemble (XGBoost, Monte Carlo, Bayesian) predicting the 2026 F1 era with 38.9% accuracy.
DATA SCIENCE • DESIGN • MOTION •DATA SCIENCE • DESIGN • MOTION •DATA SCIENCE • DESIGN • MOTION •DATA SCIENCE • DESIGN • MOTION •
Data Scientist · Designer · Builder
Basanth Periyapatna Roopa Kumar
Building things that live at the intersection of
data, design, and human experience.
Open to opportunities
About
Who I am
During my first year of engineering, two people I love survived sudden cardiac arrest. Watching a cardiologist trace every millisecond on an ECG print-out — and pinpoint the exact moment each heart faltered — was a revelation: data, interpreted with precision, saves lives.
Those late-night explorations into HMMs, SVMs, and arrhythmia detection set my compass. Coming from a family of professors, I learned that knowledge is worthless unless it's shared for social good. Today I'm a data-science practitioner who enjoys connecting the dots — ideas across disciplines, people across teams, applications across industries.
Backed by 4+ years translating complex datasets into clear actions for stakeholders in telecom, logistics, and e-commerce — I'm now pursuing an M.S. in Applied Data Science at SJSU, sharpening statistical rigor and entrepreneurial vision toward a healthcare-focused analytics startup that empowers NGOs to improve outcomes for underserved communities.
4+Years building
10+Projects shipped
4Certifications
AZ-900DP-900PL-900IIT-M · Master Data Science
PythonRC++GoPyTorchApache KafkaAWSAzureKubernetesNext.jsSQLApache SparkLangGraphScikit-learnPower BITableauDAXTypeScriptTerraformPower PlatformPythonRC++GoPyTorchApache KafkaAWSAzureKubernetesNext.jsSQLApache SparkLangGraphScikit-learnPower BITableauDAXTypeScriptTerraformPower PlatformPythonRC++GoPyTorchApache KafkaAWSAzureKubernetesNext.jsSQLApache SparkLangGraphScikit-learnPower BITableauDAXTypeScriptTerraformPower Platform
Selected Experience
Jan 2025 - Present
Jan 2025 - PresentT+16M
SJSU Research Foundation
research
Built an end-to-end AI pipeline for automated wildfire defensible space compliance using UAV imagery at 1-3 cm GSD. Engineered a Graph-RAG system over California PRC 4291 and IBHS Wildfire Prepared Home standards, representing regulatory rule chains as a NetworkX knowledge graph and injecting structured compliance context into multi-VLM reasoning chains. Applied Grounding DINO + SAM 2 for precise object detection and segmentation of trash cans, propane tanks, and vehicles. Benchmarked GPT-4o Vision, Claude Vision, and GeoChat-7B with few-shot Geo-CoT prompting as built-in XAI, producing the first peer-reviewed automated aerial parcel-level HIZ assessment system targeting California's 750K WUI properties.
"Pioneered first automated aerial combustible object detection system for wildfire HIZ compliance"
PythonGraph-RAGGrounding DINOSAM 2VLMsNetworkXUAV / Geospatial
Aug 2025
Jun 2025 – Aug 2025T+3M
NASA Ames Research Center
internship
Selected for the NASA Ames and SJSU FireSage program, a competitive research initiative at the intersection of fire ecology and geospatial remote sensing. Applied Python-driven machine learning and deep learning to production wildfire research workflows alongside active NASA scientists. Engineered a fully automated Python pipeline to retrieve and cross-validate structure built-year records across multiple public property data sources, implementing confirmation logic that measurably improved data reliability before integration into parcel-level wildfire risk datasets. Accelerated research team output by translating complex ML and deep learning concepts for non-technical collaborators, and gained hands-on HPC and distributed deep learning experience mentored directly by NASA research scientists, bridging those techniques to live project workflows.
"Engineered multi-source property intelligence pipeline validating parcel-level built-year data for NASA wildfire risk models"
PythonMachine LearningDeep LearningRemote SensingHPCData PipelinesFire Ecology
Jun 2024
Apr 2023 – Jun 2024T+15M
AtkinsRéalis
Full-Time
Developed Kafka-backed event-driven microservices on Azure Databricks with exactly-once semantics. Deployed RESTful XGBoost inference APIs via blue-green deployments, slashing infrastructure costs by $400K and reducing unplanned operational failures by 20%.
"Designed event-driven Azure tracking algorithms for 1.8M IoT signals & ML optimization"
Azure DatabricksKafkaData EngineeringXGBoostPython
Sep 2022
Sep 2021 – Sep 2022T+13M
6D Technologies
Full-Time
Built low-latency RESTful APIs in Python/C++ for production messaging architectures. Engineered stateful stream processing via Kafka, refactored PostgreSQL materialized views tuning query latency (-28%), and deployed zero-downtime microservices using GitLab CI/CD and Docker.
"Stabilized concurrent PostgreSQL databases handling 15k+ req/sec & modernized CI/CD"
PostgreSQLC++KafkaDockerCI/CD
Work
Shipped
Sequential Horizon→
02◆ LLM Multilingual Deficit
Data ·NLP / Tokenization / LLMs
LLM Multilingual Deficit
Quantifying the structural 'Token Tax' and economic inequality disadvantaging non-English languages in global LLMs.
03◆ Cartograph
Data ·React / Three.js / D3.js
Cartograph
An interactive data visualization platform that transforms raw datasets into cinematic 3D narratives.
Research & Engineering
Deep dives
AI/ML
Australian GP 2026: V1 Prediction — Russell 38.9%
Season opener under the biggest regulatory reset in F1 history. XGBoost loved Hadjar at 35.2%; Monte Carlo gave Russell 75.8% from pole physics. The ensemble resolved at 38.9% — and documented three V1 failures for V2.
AI/ML
Japanese GP 2026: Antonelli on Pole at Suzuka
Post-Australia recalibration tightened the model's variance. Antonelli's first career pole at 19, Verstappen buried in P14 from an ERS error. Mercedes 1-2 probability: 84.5%. The Suzuka circuit amplifies what qualifying says.
AI/ML
Chinese GP 2026: V2 — Sprint as Tier 0 Data
First sprint weekend of 2026. V2 rebuilt with live sprint race results as the highest-weight input layer. Hamilton re-rated to 21.2% after his V1 underestimation. The fp1_to_quali_divergence feature corrected a systematic blind spot.
Research
The Multilingual Deficit: Tokenization Inequality in LLMs
Kannada requires 3-5× more tokens than English for equivalent semantic content. Quantifying the Token Tax, the $262K-$365K annual cost premium for Hindi API services, and the three architectural bottlenecks that cascade from tokenization.
AI/Health
Cognivi — Stroke Detection
A privacy-first, multimodal AI system built at TreeHacks 2026. MediaPipe arm-drift + facial asymmetry detection fused with a PubMed-grounded RAG pipeline feeding Claude 3 for neurologic speech grading — all from a 60-second smartphone video.
Academic Work
Research & Papers
Peer-ReviewedBig Data · Healthcare2025
A Big Data Pipeline Approach for Predicting Real-Time Pandemic Hospitalization Risk
Vishnu S. Pendyala, Mayank Kapadia, Basanth Periyapatna Roopa Kumar, Manav Anandani, Nischitha Nagendran
A dual big-data pipeline for real-time pandemic triage combining a streaming epidemiological risk prediction system (XGBoost + Bloom filter pre-screening on 3M+ CDC COVID-19 records) with a chest X-ray classification pipeline (EfficientNet-B0 + Grad-CAM). A lightweight GPT-based reasoning layer generates auditable ALERT/FLAG/LOG triage comments. CTGAN validates streaming robustness under synthetic load. Provides scalable, explainable, near-real-time decision support for public health readiness.
XGBoost Minority F1
0.76
Chest X-Ray Accuracy
99.5%
CDC Records Processed
3M+
Generative AIHealthcare2025
A Multimodal Generative AI Framework for Cancer Pathology Classification
Mayank Kapadia, Basanth Periyapatna Roopa Kumar, Nischitha Nagendran
A multimodal generative AI framework combining histopathology image classification, clinical note classification via ClinicalBERT with SFT/TAPT/LoRA fine-tuning, and prompt-driven clinical captioning using RAG. Establishes baselines on PatchCamelyon (262K images) and curates TCGA BRCA clinical notes into a balanced dataset of 2,380 reports for unified multimodal cancer diagnostics.
ViT-Base/16 ROC-AUC
0.9601
ClinicalBERT LoRA F1
0.94
PCam Test Images
32,768
Deep LearningRemote Sensing2025
FireSense: Multispectral Fire Detection with Channel Attention and Probabilistic Calibration
Basanth Periyapatna Roopa Kumar, Nischitha Nagendran, Nandhakumar Apparsamy, Nitya Rondla
A deep learning framework for automated wildfire detection in Landsat-8 satellite imagery. Combines ResNet34-UNet with Convolutional Block Attention Modules (CBAM) across all 10 spectral bands, introduces soft labels from multi-annotator consensus to capture boundary uncertainty, and applies temperature scaling for calibrated probability estimates ready for operational deployment.
Mean IoU
69.6%
vs. Classical Baselines
+45.6%
ECE Reduction
86.1%
Knowledge GraphsML2025
Knowledge Graph-Enhanced Traffic Optimization System for the Bay Area
Anshu Reddy Dhamana, Basanth Periyapatna Roopa Kumar, Manav Rajesh Anandani, Nischitha Nagendran, Nitya Rondla, Srithareddy Devireddy, Vinuthna Papana
A knowledge graph-based traffic optimization system for the San Francisco Bay Area fusing traffic flow, weather, incidents, and events into a unified semantic graph. Node2vec embeddings power a stacking ensemble achieving 92% prediction accuracy with the connects_to relationship, delivering actionable insights into weekday vs. weekend traffic patterns for city planning and commuter support.
Prediction Accuracy
92%
Graph Relationship Types
6
Graph Nodes
550+
CapstoneLLM Fine-Tuning · RAG2026
Professor in a Box: Personalized CS Tutoring via Fine-Tuned LLMs and Retrieval-Augmented Generation
Basanth Periyapatna Roopa Kumar, SJSU MSDA Team 2
An end-to-end AI tutoring system that fine-tunes Mistral-7B and DeepSeek-R1-7B using QLoRA on curated CS Q&A corpora and serves them via a production RAG pipeline. Students upload study materials; the system retrieves semantically relevant context using Gemini embeddings and ChromaDB, then streams grounded answers token-by-token from LoRA-adapted models running on serverless A10G GPUs. A three-tier fallback chain (OpenAI, DeepSeek, Gemini) ensures continuous availability.
Fine-Tuned Models
2
VRAM Footprint
~4 GB
RAG Retrieval Top-K
5
Library
What I'm consuming
Books read & screens watched in 2025. Hover to read my take.
★★★★★
The Design of Everyday Things
Don Norman
Design
★★★★★
Sapiens
Yuval Noah Harari
History
★★★★★
The Creative Act
Rick Rubin
Creative
★★★★★
Zero to One
Peter Thiel
Business
★★★★★
Thinking, Fast and Slow
Daniel Kahneman
Psychology
★★★★★
The Almanack of Naval Ravikant
Eric Jorgenson
Philosophy
★★★★★
Show Your Work!
Austin Kleon
Creative[M·06] SIGNAL: STRONG
Content I'm creating
Quotes
Words I live by
Vision
Thepeoplewhoarecrazyenoughtothinktheycanchangetheworldaretheoneswhodo.
Steve Jobs
01 / 05
Design
Simplicityistheultimatesophistication.
Leonardo da Vinci
02 / 05
Growth
Youdon'thavetobegreattostart,butyouhavetostarttobegreat.
Zig Ziglar
03 / 05
Agency
Thebestwaytopredictthefutureistocreateit.
Peter Drucker
04 / 05
Life
Stayhungry.Stayfoolish.
Stewart Brand
Whole Earth Catalog
05 / 05
Get in touch
Let'smakesomething
Have a project in mind? A collaboration idea? Or just want to say hi?
My inbox is always open.
Or find me on
Open to opportunities