Parth Maniar

I’m Parth Maniar — a Senior Software Engineer building AWS data platforms and LLM-powered products for neuroscience and healthcare teams. Based in San Jose, CA. M.S. in Data Science (University at Buffalo, SUNY).

On AWS, I build end-to-end systems—ingestion, transformation, analytics, and ML—with observability, security, data quality, and cost control baked in from day one. Recently, I’ve shipped RAG pipelines, evaluation harnesses, and workflow automation that turn clinical, imaging, and operational data into auditable, decision-ready insights.

I’m a pragmatic engineer who values ownership, fast feedback loops, and clear, lightweight documentation. I write about data engineering, LLM/RAG systems, and operational reliability, and contribute to open source. If you’re building on AWS or shipping LLMs to production, I’d love to connect—email me, message me on LinkedIn, or book a 30-minute call.

Dashboards don’t fail loudly when upstream data drifts. Data contracts make breakage explicit so teams can ship faster with less panic.

Data Contracts Before Dashboards: Stop Shipping Silent Data Bugs

You don’t need a feature store on day one, but you do need strict consistency between training and serving. Ignore that and model performance will rot in production.

Feature Stores Are Optional. Feature Consistency Is Not.

RAG isn’t “set and forget.” If you don’t define acceptable failure rates and monitor retrieval quality, your assistant will eventually confidently ship nonsense.

Why Your RAG System Needs a Failure Budget

Idempotency is the simplest idea that saves you from duplicate charges, double emails, and messy retries. Here's how to design it in real systems.

Idempotency Is Your System's Seatbelt: Building Reliable APIs & Job Queues

The prototype is easy. The on-call rotation is not. Here’s a practical, story-driven checklist for shipping LLM features that don’t collapse under real users.

The LLM Shipping Checklist: From Demo to Production (Without Getting Burned)

Exploring how Retrieval Augmented Generation helps large language models access external knowledge, reduce hallucinations, and deliver more reliable responses by knowing when to retrieve rather than generate.

Retrieval Augmented Generation: Building LLMs That Know Their Limits

Explore the critical difference between AI Agents and Agentic AI with real-world examples from Waymo, Tesla, and Google. Learn how autonomous systems are reshaping technology.

AI Agents vs Agentic AI: The Critical Distinction Explained

Discovering the Power of Dimensionality Reduction - Unveiling patterns and visualizing data with PCA and t-SNE.

Understanding PCA and t-SNE: A Comparative Analysis for Dimensionality Reduction

Hyperparameter tuning in machine learning, including methods such as grid search, random search, and Bayesian optimization and providing code snippets to demonstrate how to use these techniques in practice.

Cracking the Code: Mastering Hyperparameter Tuning for Optimal Machine Learning Performance

Model interpretability in machine learning, including the use of LIME and SHAP techniques

Unlocking the Secrets of Machine Learning: Techniques for Interpreting Complex Models

Utilize the power of Facebook Prophet to identify anomalies in your time series data and safeguard your data from potential issues with the help of Python.

Uncovering the Unusual: Using Facebook Prophet for Anomaly Detection in Time Series Data

Practical guide to identifying and removing outliers in Python using Z-score, IQR, and visualization methods. Boost your data quality with these data cleaning techniques.

Out with the Outliers: How to Remove Outliers in Python

Debugging a model or data science project involves following a systematic approach to identify and fix issues at each stage of the pipeline, starting with the data loaders and ending with the outputs.

Debugging a model or any data science project

Regression is a supervised learning algorithm used to predict continuous or real-valued outputs for given inputs, and is widely applied in a variety of fields to make accurate predictions.

The Ins and Outs of Regression: Types, Training, and Evaluation

Find the missing data percentage and count in pandas

Missing Data Percentage and Count in pandas

A complete guide to generating Spotify Refresh Tokens for your applications. Learn OAuth2 flows and token management with step-by-step instructions.

How to get the Spotify Refresh Token

Apple announced that it would begin checking content uploaded to iCloud Photos, against a list of known CSAM material. This is fine on its own, however the system could easily be abused.

Apple doesn't care about privacy, and it's showing.

The EU announced plans to force all devices – including iPhones – to use a common charging port.

Apple and the case for competition.

Streamlit app that flags potentially fraudulent credit card transactions using a Random Forest model.

Credit Card Fraud Detection

Model-driven screening of Buffalo, NY properties for investment using public assessment roll data.

Property Investment Assessment

A gamified waste-segregation app that classifies garbage (organic vs non-organic, wet vs dry) using a deep-learning model trained on a custom dataset.

SpotGarbage

Unified AI platform for agriculture forecasting (yield, weather, price, and warehousing) using 10k+ models trained on government data across 55+ crops and 343 Indian districts.

Unified System for Agriculture Prediction using AI (USAPA)

Virtual queuing for essential services during COVID-19 — reduce crowding with predicted wait times and e-tickets.

Covid SafeQ

SARIMAX time-series forecasting of recovered COVID-19 cases in India (train: Mar–Aug 2020; forecast: late Aug–Sep), with predicted vs. actual comparisons.

Recovered Cases Forecasting (COVID-19, India)

End-to-end quarantine management: symptom check-ins, geofence breach alerts, and COVID hotspot prediction.

QuaranTime App

React portal for the KJSIEIT IT Department to share webinar resources, recordings, and certificates.

KJSIEIT IT Department — Resources Distribution Site

CNN-based waste image classifier that suggests the correct recycling category from a photo.

Smart Waste Segregator (Classifier)

A smart-lock prototype with app, voice, and keypad unlock plus audit logging.

Ingenious Lock System

Flask + Selenium web app that downloads SlideShare decks from a URL for offline viewing (use responsibly).

SlideShare Downloader 📥

Android app that predicts whether outdoor conditions are favorable using a Decision Tree classifier on basic weather features.

Play Or Not App

AI-assisted capacity planning to prioritize public toilet locations and capacity, with a government portal, citizen app, and IoT hygiene monitoring.

Hi, I'm Parth Maniar

Featured projects

Unified System for Agriculture Prediction using AI (USAPA)

QuaranTime App

Enhancing Sanitation By Capacity Planning of Public Toilets

Latest posts

Data Contracts Before Dashboards: Stop Shipping Silent Data Bugs

Feature Stores Are Optional. Feature Consistency Is Not.

Why Your RAG System Needs a Failure Budget