Aman Saini
Aman headshot

Machine Learning Engineer

Seasoned ML Engineer specializing in LLM post-training, fine-tuning, retrieval, LLM evaluation, Deep Learning, Natural Language Generation, Natural Language Understanding

saini.aman.personal@gmail.com +1 (360)842-6055 Vancouver, BC, Canada
LinkedIn Google Scholar

Experience

Apple — Machine Learning Engineer (MLE)

Apr 2025 – Present | Vancouver, BC, Canada
  • Building cutting-edge Apple Intelligence features by training and optimizing GenAI models using advanced post-training methods such as DPO and Online Reinforcement Learning.
  • Designing Verifiable Rewards pipelines with Rubrics and Reward Models to improve alignment and model quality.
  • Developing scalable evaluation and training frameworks for high-quality generative AI experiences.

Grammarly — Staff ML Engineer

Oct 2023 – Apr 2025 | Vancouver, BC, Canada
  • Strategic Research org advancing LLM capabilities in the writing space; previously Core Product building Generative AI features.
  • Fine-tuned LLMs (Llama, GPT) for fine-grained writing assistance; instruction-tuned Llama 3.1 (70B/8B) teacher models.
  • Built evaluation pipelines using LLM-as-a-Judge; scaled synthetic data generation for internal LLMs.
  • Published multi-task Ukrainian text-editing work; released models/datasets on Hugging Face.
  • Created a Python library for automatic prompt optimization with target reward metrics (EMNLP 2023 style approach).

Twitter, Inc. — Senior ML / NLP Engineer

May 2021 – Jan 2023 | Seattle, WA, USA
  • NLP Signals (Cortex): end-to-end NLP models & signals used across Home Timeline, Notifications, and Trends.
  • Entity Linking: Encoder-only Transformers for detection, candidate generation, and ranking; dataset released at NeurIPS 2022; signals used in production.
  • NER: Replaced Bi-LSTM with multilingual BERT using subword masking, token features, weak labels, and distillation → +13.5% F1.
  • Tweet Representation: Multi-task model for language, topical, and engagement prediction; published in DL4SR 2022.
  • Improved notification seed Tweet quality; led NER adoption and ML code guidelines.

Microsoft — Senior ML Engineer

Jan 2016 – Apr 2021 | Seattle, WA, USA
  • Bing Ads (AI+Research): improved ad experience and launched new ad products to drive clicks.
  • Dynamic Search Ads: Landing-page extraction & Transformer-based generation → +1.5% RPM (US daily), 2%+ RPM international.
  • Web Page Similarity Graph: TwinBERT/USE embeddings; scalable ANN (HNSW/NSG) for next-link prediction.
  • Dynamic links with ads: DeepXML categories + Ads RoBERTa scoring → +1.5% CTR (US).
  • Landing page summarization: BERTSUM for extractive summaries used in ad title assets.
  • Ad Creative infrastructure: Algorithms integration, editorial service, Azure, caching; latency tooling adopted across services.

Microsoft India — Software Engineer

Jul 2013 – Dec 2015 | Bangalore, India
  • Rich Ads experience (RnR): online infra + rich ad experiences for non-US markets.
  • Ads-Composition: Unified, parallelized pre-serve pipeline for ads & decorations → lower latency, faster experimentation.
  • Related Product Annotation: Surfaced relevant products for retail queries → higher clicks on text ads/product links.
  • Latency tooling: Utilities to analyze critical latency paths, reused across services.

Publications

Spivavtor: An Instruction-Tuned Ukrainian Text Editing Model

Aman Saini, Artem Chernodub, Vipul Raheja, Vivek Kulkarni

arXiv 2024

We introduce Spivavtor, a Ukrainian-focused instruction-tuned text-editing model and dataset. Spivavtor adapts the CoEdIT framework to Ukrainian and provides models and data for instruction-based text rewriting/editing. We describe the dataset construction, model training, and evaluation results.

TweetNERD — End-to-End Entity Linking Benchmark for Tweets

Shubhanshu Mishra, Aman Saini, Raheleh Makki, Sneha Mehta, Aria Haghighi, Ali Mollahosseini

NeurIPS — Datasets & Benchmarks Track, 2022

We present TweetNERD, a large-scale benchmark for entity extraction and linking on Twitter. The dataset spans 2010–2021 with three tasks: NER, Entity Linking with gold spans, and End-to-End Entity Linking. We provide strong baselines and analyze in-domain and out-of-domain performance.

APIO: Automatic Prompt Induction and Optimization for Grammatical Error Correction and Text Simplification

Artem Chernodub, Aman Saini, Yejin Huh, Vivek Kulkarni, Vipul Raheja

arXiv 2025

We propose APIO, a two-stage approach that first induces task-specific prompts from examples and then optimizes them for text editing tasks. On Text Simplification, APIO attains strong SARI scores on ASSET-Test. We outline the induction/optimization pipeline and report evaluations across simplification and grammatical error correction.

Internships

2011 – 2013

Amazon — Software Engineer Intern (Pricing)

Jan – Jun 2013 | Bangalore, India

  • Worked for the Amazon Pricing team to predict the shipping price for products sold on amazon.com.
  • Implemented the end-to-end Pricing Attribute Prediction Service using Amazon Simple Workflow Service (SWF) to automate the ML model building process.

University of Victoria — Research Intern

May – Jul 2012 | Victoria, Canada

  • Replaced Java with the bidirectional language Boomerang in a high-confidence medical data device (MirthConnect) to ensure correctness of transformations and guarantee well‑behaved channels.

Indian Institute of Remote Sensing (IIRS) — Research Intern

May – Jul 2011 | Dehradun, India

  • Worked with the Weather Research and Forecasting (WRF) Modeling team to deploy WRF models on Linux systems, recreating past events for correctness and enabling real‑time weather prediction.

Education

Birla Institute of Technology and Science (BITS) Pilani

B.E. in Computer Science • CGPA 9.54/10

2009 – 2013

Contact

Email
saini.aman.personal@gmail.com
Phone
+1 (360)842-6055
Location
Vancouver, BC, Canada