~ a new quest begins ~
ParthKatlana
Building production AI systems that actually work when it matters most.
About
Character Sheet
Parth Katlana
Data Scientist
I'm an engineer who believes the most complex data should serve the most critical missions. Currently at NCICS & NOAA, I build scalable, cloud-native architectures to make sense of our world's climate data. With a Master's in Computer Science from NC State and a deep love for MLOps, I turn distributed systems into engines for real-world impact — I don't just build models, I build the infrastructure that ensures they work when it matters most.
Skills
Connect
Loves building side projects · Loves travelling · Loves coffee
Achievements Unlocked
NASA Grant
Hallucination Detector
Papers Published
GPU Cost Cut
Certifications
Users Scaled (A/B)
Journey So Far
Experience
Every stop, a story. Every role, a relic.
scroll to sail the journey
NCICScurrent
Data Scientist
Jan 2025 - Present
- ▸Secured $500k NASA grant by designing a novel forecasting prototype that reduced RMSE by 17.4% vs CNN baselines.
- ▸Designed and shipped a production agentic AI system (LangChain, RAG, LLMs) that autonomously plans multi-step workflows over 12-25 TB/mo of satellite data, cutting time-to-insight from ~3 hrs to 20 min.
- ▸Built an evolutionary agentic system where LLM agents propose model-architecture mutations, orchestrate training & evaluation via Step Functions, and select optimal designs through multi-objective fitness.
- ▸Productionized deep learning models with FSDP and CUDA distributed training, reducing GPU training cost by 40% and tile inference latency to 250 ms.
Laboratory of Analytical Sciences
Data Scientist
May 2023 - Aug 2024
- ▸Orchestrated an LLM ensemble (Mistral, Llama-2, GPT-4) with a ranking layer that transforms unstructured text into structured reports, reducing review time by 50%.
- ▸Built an LLM evaluation framework with a metric-ensemble hallucination detector (0.91 Pearson correlation) to validate factual consistency, fairness, and robustness — published 2 papers.
- ▸Fine-tuned in-house CLIP and YOLOv8 models for domain-specific detection, increasing mAP50-95 by 15%.
- ▸Developed classification models (XGBoost, Logistic Regression) with Bayesian optimization for risk assessment.
Kion Technologies
Data Scientist
Aug 2019 - Jan 2022
- ▸Scaled an A/B testing infrastructure to 10M+ daily active users, reducing system latency by 40% through ML model integration.
- ▸Performed credit risk profiling and customer segmentation (logistic regression, K-Means) on structured and behavioral metadata.
- ▸Optimized ETL pipelines across Snowflake and SQL Server, building regression and multivariate models enabling real-time KPI visibility.
- ▸Built data visualizations (Power BI, Tableau, Matplotlib) that increased marketing campaign effectiveness by 25%.
Projects
Quest Log
~ legendary encounters, each worth remembering ~
I.Agentic AI System for Climate Data
Production agentic system (LangChain, RAG, LLMs) that autonomously plans multi-step workflows — retrieving context, calling tools, running analyses, generating structured reports over 12-25 TB/mo of satellite data. Cut time-to-insight from ~3 hours to 20 minutes.
Developed an ensemble approach for detecting hallucinations in abstractive text summarization. Combined unsupervised metrics to demonstrate LLM-based methods are more effective at identifying hallucinations, achieving 0.91 Pearson Correlation — surpassing the previous state-of-the-art.
III.Evolutionary Neural Architecture Search
Built an evolutionary agentic system where LLM agents propose model-architecture mutations, orchestrate training & evaluation via Step Functions, and select optimal designs through multi-objective fitness with built-in failure handling.
IV.Agentic Credit-Risk Underwriting
Multi-agent system (LangGraph, LLMs) automating PD scoring (XGBoost), RAG-based Basel III/IV regulatory validation, and SHAP-driven decision explanations for transparent, compliant credit decisions.
Side Quests
~ bonus adventures worth mentioning ~
Traffic Monitoring System
Computer Vision system using HOG/SVM classifiers to reduce wait times for cars and pedestrians in high-traffic zones.
Find My Roomie
Web app for NC State students to find roommates, built with Django, React, and PostgreSQL following software engineering best practices.
Music Genre Classification
Classifier working with frequencies and amplitude to categorize sound clips, with a recommender system using SVMs, Random Forests, and Neural Networks.
Treasure Found
Let's Connect
take what you need, adventurer.