Resume · experience, projects, education

Where I've worked

Ads targeting and user modeling at TikTok, research engineering at Stanford and Microsoft Research, and ML infrastructure at Amazon. Plus three research projects from Stanford.

Viewer 1 is into
The 3 ads this viewer sees
    1B ads→1,000→50→3 shown
    An ad's score is its bid times the chance this viewer clicks it. Pick one interest or several.
    1. TikTok

      Current

      Machine Learning Engineer

      Jul 2024 to Present · 2 yrs 5 mos

      • Owned the revamp of TikTok ads gender targeting across the US, EU and the rest of the world, covering 2B+ monthly users.
      • Replaced a stale US gender MLP with XGBoost trained on identity-matched third-party labels and census-calibrated thresholds, raising F1 from about 83% to about 90% and advertiser value by 0.44% (p < 0.01) at full launch.
      • Launched the EU gender model using computer-vision, social-graph and viewing signals: male recall up 15 points, coverage from 85% to 98.5%, and the male/female split corrected from 26/74 to 45/55.
      • Designed an LLM user-modeling system that maps ads into about 450K semantic clusters and fine-tunes Qwen2.5 and Yi (up to 7B) on six-month behavior sequences of 91M users into 64-dimensional lookalike embeddings.
      • Cut LLM fine-tuning time 150x, from 100+ hours to 40 minutes, with dynamic padding and PyTorch compilation.
      • Lifted weighted AUC 8.5% on gaming-ad purchase-value models with Huber and relative-absolute-deviation losses.
      • Ads targeting
      • XGBoost
      • LLM fine-tuning
      • User embeddings
      • PyTorch
    2. Stanford Online Deliberation Platform

      Senior Research Software Engineer

      Sep 2023 to Jun 2024 · 10 mos

      • Built an LLM pipeline with GPT-4 that summarized 800 deliberation sessions and 12,000 hours of group discussion.
      • Designed the real-time video architecture (Amazon Chime SDK, WebRTC) for a platform with 50,000+ participants.
      • GPT-4
      • WebRTC
      • Amazon Chime SDK
    3. Microsoft Research

      Research Software Engineer Intern

      Jun 2023 to Sep 2023 · 4 mos

      • Optimized Azure VM allocation by analyzing VM lifetimes, reducing fragmentation by 3.5%.
      • Developed a real-time dashboard to analyze VM metrics across 100+ Azure availability zones.
      • Azure
      • Data analysis
      • Dashboards
    4. Stanford Reading & Dyslexia Research

      Research Software Engineer

      Sep 2022 to Sep 2023 · 1 yr 1 mo

      • Built a dyslexia screening platform in Vue.js and Node.js used to assess 8,000+ students.
      • Automated score-report generation on Google Cloud for 600+ teachers across 50 schools.
      • Re-architected serverless cloud functions, cutting pipeline runtime from 4 hours to 30 minutes.
      • Vue.js
      • Node.js
      • Google Cloud
    5. Amazon

      Software Development Engineer

      Jan 2021 to Sep 2022 · 1 yr 9 mos

      • Designed an ML-as-a-service platform with modules for scaling, model evaluation and retraining.
      • Architected model retraining across 19 Amazon marketplaces using AWS Step Functions.
      • Cut data-processing time from 16 hours to under 30 minutes with optimized C++ operations.
      • Scaled concurrent AWS Lambda executions 175x and added alarms for failures and timeouts.
      • Conducted 30+ on-site interviews for intern and graduate roles; mentored 50+ university students.
      • MLOps
      • AWS
      • C++

    Projects

    Research projects from my time at Stanford, in generative models, multimodal dialogue and NLP.

    SOTABeat state of the art on SSIM and FID

    Facetime: Face Prediction with Diffusion Models

    Sep to Dec 2023

    Generates the face of a child from photos of both parents, plus aged versions at future intervals. Beat state-of-the-art models on both SSIM and FID.

    • Diffusion models
    • Computer vision
    +7.58%Improvement over the baseline

    npcGPT: Multimodal Dialogue for Social Robots

    Apr to Jun 2023

    Combines visual context with environment data to generate human-like conversation and behavior for social robots, improving on the baseline by 7.58%.

    • Multimodal
    • LLMs
    3rd / 163Teams, at 76.9% accuracy

    Multi-task BERT

    Jan to Mar 2023

    Fine-tuned BERT for sentiment analysis, paraphrase detection and semantic similarity using contrastive learning with multiple negatives. Ranked 3rd of 163 teams at 76.9% accuracy.

    • NLP
    • BERT
    • Contrastive learning

    Skills

    Recommendation
    Two-tower models, Multi-task ranking, Mixture of experts, Lookalike modeling, User embeddings, GNNs
    LLMs
    Qwen2.5, Yi, Llama, Full fine-tuning, LoRA, FlashAttention, Prompt engineering
    Languages and frameworks
    Python, C++, SQL, Java, PyTorch, TensorFlow, Hugging Face, XGBoost
    Data and cloud
    PySpark, Spark SQL, Hive, HDFS, Kafka, AWS, GCP
    Experimentation
    A/B testing, Offline evaluation, Model calibration, AUC, Precision and recall at K, MRR

    Education

    1. 2022 to 2024

      Stanford University

      M.S. Computer Science, AI and Software Theory · GPA 4.0 · Distinction in Research

    2. 2016 to 2020

      BITS Pilani

      B.E. Computer Science · CGPA 9.82 / 10 · merit scholarship all 8 semesters

    Competitive programming

    1. 2020

      Google Hash Code

      1st in UAE, 689th of 10,724 teams

    2. 2020

      Google Kick Start

      291st worldwide

    3. 2021

      Google Code Jam

      Round 2 qualifier

    4. 2018, 2019

      Gulf Programming Contest

      1st place