California, USA

Hi, I'm Darshil. I build ML systems

I'm an SDE II on Amazon Ads, where I build large-scale ad-tech systems — brand-safety integrations, high-throughput data pipelines, and microservices that run at Prime Day scale. My background spans machine learning, data science, and cloud computing, and I like turning messy, large-scale data into systems that are fast, reliable, and genuinely useful.

See my experience Get in touch

About

Engineer at the intersection of data, ML & the cloud.

I'm an SDE II on Amazon Ads, building large-scale ad-tech systems — brand-safety integrations on Amazon DSP, Spark pipelines processing terabytes a day, and microservices that hold up at Prime Day scale. My background spans machine learning, data science, and cloud computing, and I care about building software that's fast, reliable, and actually used.

Outside of shipping code, I like learning new things and trading ideas with people who are just as curious. Based in California, USA.

Open Source

Experience & Education

Where I've worked & studied.

Work

  1. Jun 2022 — Present

    Software Development Engineer II

    Amazon Ads

    • Integrated industry-leading brand-safety products (DoubleVerify, IAS) into Amazon DSP; designed cache-miss handling and offline cache population in the Ad Exchange server, sustaining 300K+ QPS during Prime Day.
    • Architected Spark pipelines on Amazon EMR orchestrated by AWS Step Functions, processing 3TB+ of operational data daily and cutting complex-log query time ~30% to speed up debugging and incident response.
    • Designed an automated publisher-onboarding system for Amazon DSP that reduced onboarding from 6+ weeks to under 2 hours — a 99%+ reduction in processing time.
    • Led 2 engineers to build a search-query ad-blocking system with a pipeline automating 60K+ keyword annotations, blocking ~2% of ad traffic and increasing advertiser brand trust.
    • Led 2 engineers to design a high-volume microservice collecting user-action events across Amazon Advertising (with PII compliance) plus its visualization UI, driving a year-over-year reduction of 5,000+ customer support contacts.
    • Mentored 1 intern (converted to full-time) and 2 SDEs through technical ramp-up and growth.
  2. Jul 2019 — Jul 2022

    Senior Software Engineer

    Oracle

    • Designed and shipped features for Oracle's CX Cloud CRM platform, adopting modern conversational-UI patterns across a full-stack JavaScript and microservice codebase for enterprise customers.
    • Architected a central navigation framework and event-based data-loading system for Oracle CX Cloud, simplifying complex workflows and improving client-side state management for 100K+ daily users.
    • Built a reusable UI component library that let internal teams and external customers extend the platform with minimal custom development, reducing customization effort for enterprise implementations.
    • Drove accessibility initiatives using assistive technologies (JAWS) to make the platform usable across a wide range of abilities.
  3. Jul 2018 — Aug 2018

    Software Engineer Intern

    Cox Automotive Inc.

    • Built ML models (Logistic Regression, XGBoost) to predict dealer contract probability at 94.5% accuracy, helping dealers prioritize their time, and used a correlation matrix to identify the most influential parameters for feature engineering.
    • Refactored the core product application to React-based components, improving code modularity and maintainability.
  4. Aug 2015 — Jul 2017

    Systems Engineer

    Tata Consultancy Services

    • Delivered software quality and automation across client engagements for Dell EMC and the Kenya Revenue Authority. At Dell EMC, owned full test coverage for the SYMMETRIX (DMX/VMAX) remote-replication disaster-recovery product and was among the first to introduce Python automation alongside Linux/Windows scripts.
    • For a Hibernate/Spring Java web app, built Selenium test automation and mapped scenarios to business requirements for complete coverage, reporting bi-weekly release status and risk plans to clients.

Education

  1. 2017 — 2019

    M.S. in Computer Science

    New York University

    GPA 3.7/4.0 · Machine Learning, AI, Cloud Computing, Big Data, Algorithms

  2. 2011 — 2015

    B.E. in Computer Engineering

    Gujarat Technological University

    GPA 8.33/10 · Ranked 3rd in class · Distributed & Parallel Computing, Compilers

Academic Projects

Earlier ML, data & cloud work from grad school.

Shopper Sentiment Analysis 01
Cloud · Computer Vision

Shopper Sentiment Analysis

Actionable retail insights derived from live in-store video streams.

Problem
Brick-and-mortar retailers have rich in-store behavior happening in front of cameras every day, but almost none of it is captured as data. Store owners couldn't answer basic questions: which aisles draw the most traffic, who their shoppers are, or how customers actually feel while browsing.
Approach
Built a pipeline that ingests live store video and applies managed ML vision services to detect foot-traffic hotspots, segment shoppers by demographic (approximate age, gender), and infer sentiment from facial expression — turning raw footage into a queryable stream of behavioral signals.
Impact
  • Surfaced high- vs. low-traffic zones to inform product placement
  • Demographic segmentation of store traffic in near real time
  • TODO: add a hard metric (e.g. frames/sec processed, accuracy, cost per store/day)
  • AWS Rekognition
  • AWS Lambda
  • S3
  • Python
Restaurant Chatbot & Recommendation System 02
Cloud · Serverless

Restaurant Chatbot & Recommendation System

A serverless conversational concierge that recommends restaurants by location, cuisine & rating.

Problem
Finding the right restaurant means juggling location, cuisine, date, and ratings across multiple apps. There was no single conversational interface that could take a natural request and return a tailored recommendation — delivered the way the user prefers.
Approach
Designed a fully serverless dining concierge. A web front end talks to a chatbot backed by Lambda; restaurant data scraped from Yelp & Google APIs lands in DynamoDB; incoming requests queue through SQS and recommendations are retrieved via Elasticsearch over an ML-trained dataset, then delivered by email or SMS.
Impact
  • End-to-end serverless architecture — no servers to manage, scales to zero
  • Recommendations delivered via the user's channel of choice (email/SMS)
  • TODO: add a metric (e.g. # restaurants indexed, median response latency)
  • AWS Lambda
  • DynamoDB
  • Elasticsearch
  • SQS / SNS / SES
  • Cognito
  • Node.js
Smart Photo Album 03
Cloud · NLP

Smart Photo Album

A photo album you can search by voice or text in plain language.

Problem
Photo libraries grow faster than anyone can organize them. Finding "the photo of a dog at the beach" usually means scrolling endlessly, because albums are searchable by filename and date — not by what's actually in the picture.
Approach
Built a photo album web app searchable through natural language via both text and voice. An intelligent search layer indexes photos by the people, objects, actions, and landmarks they contain, so a spoken or typed query maps directly to matching images.
Impact
  • Natural-language search over image content, not just metadata
  • Voice and text input both supported
  • TODO: add a metric (e.g. label accuracy, index size, query latency)
  • AWS Rekognition
  • Elasticsearch
  • Lex
  • Lambda
  • S3
Yelp Dataset Analysis 04
Big Data

Yelp Dataset Analysis

Big-data analytics over millions of Yelp reviews to compare distributed programming paradigms.

Problem
The Yelp open dataset is large enough that single-machine tools choke on it — 4.1M reviews, 947K tips, 1M users, 144K businesses. The challenge was both analytical (what can this data tell us?) and architectural (which distributed paradigm handles it best?).
Approach
Ran the same analyses across Spark, Spark SQL, and Pig Latin on Hadoop to directly compare the programming paradigms, then visualized the results — distributions, graphs, and maps — in Zeppelin and Tableau.
Impact
  • Processed a multi-million-record dataset across three distributed engines
  • Side-by-side comparison of Spark / Spark SQL / Pig for the same workloads
  • Statistical analysis and geo-visualizations of business & review trends
  • Apache Spark
  • Spark SQL
  • Pig
  • Hadoop
  • Zeppelin
  • Tableau
Pac-Man AI Agent 05
Artificial Intelligence

Pac-Man AI Agent

A self-playing Pac-Man driven by six classical search & optimization algorithms.

Problem
Pac-Man is a deceptively rich planning problem: a large state space, adversarial ghosts, and the need to balance exploration against reward. It's an ideal testbed for comparing how different AI search strategies behave under the same constraints.
Approach
Implemented an AI agent that plays Pac-Man using six algorithms — DFS, BFS, A*, Hill Climbing, a Genetic Algorithm, and Monte Carlo Tree Search — so their behavior, optimality, and runtime could be compared on identical game states.
Impact
  • Six search/optimization strategies implemented and benchmarked head-to-head
  • Clear illustration of trade-offs between optimality and compute cost
  • TODO: add a metric (e.g. avg. score or win-rate per algorithm)
  • Python
  • Search Algorithms
  • Genetic Algorithms
  • Monte Carlo Tree Search

Contact

Let's build
something together.

I'm always up for a good conversation — about ML, distributed systems, or what you're working on. The fastest way to reach me is email.

darshil.patel.2810@gmail.com