In early 2023, I decided to resign from Tinder and explore the AI startup landscape by working at different startups. During this exploration, I worked at a few startups and got hands-on experience on how Large Language Models can be leveraged to build products.

• Features:
  • Record Matching between Salesforce records and incomplete leads using a cascade of GPT 3.5 Turbo and GPT 4 using Langchain for async batching and calling the API.
  • Rate professionals for job placements based on their work experience using GPT 4.
• Fine-tuning: Fine-tuned open-source LLMs with LoRA, replacing costly proprietary systems to improve cost and stability.
• Inference: I advised infrastructure engineers, did node local benchmarks for LLM inference engines like vLLM, and evaluated services like Together.ai and Fireworks.ai.
• Knowledge: Leadership: Stayed abreast of the latest SOTA transformer architectures, such as KV Cache, Multi-Query Attention, and RoPE, ensuring the dissemination of this knowledge.
• Evaluation: Built a UI tool using Streamlit and ideas from different evaluation repositories like LLMJudge to build a tool to compare the performance of different LLMs on a given task.

Technical Lead for a Machine Learning Engineers & Risk Analysts team. I started the group and grew it to 9 people. This team owns detection models/algorithms for identifying Trust & Safety violations.

• Responsible for hiring, mentoring, and team charter.
• Organized Risk Analysts under the Escalations and Early Warning team. Collaborated with the team to invent KPIs for the org. This team optimizes detection strategies using existing signals and provides feedback highlighting detection gaps.
• Removed data scarcity by establishing best practices for leveraging moderation logs for training and evaluation datasets.
• Established modeling best practices. Transformers for text classification by pre-training on Tinder’s unique data; Text pre-processing pipelines to handle adversarial attacks; Metadata embeddings to improve multi-lingual performance; ConvNeXT family of models for effective and efficient image classifiers.
• Enabled real-time model inference. TFLite and Quantization for CPU-based models; TensorRT and Triton for GPU-based models; Collaborated with Infra team to establish requirements for the K8-based model endpoint framework.

Technical Lead for a Machine Learning Engineers & Risk Analysts team. I started the group and grew it to 9 people. This team owns detection models/algorithms for identifying Trust & Safety violations.

• Responsible for hiring, mentoring, and team charter.
• Organized Risk Analysts under the Escalations and Early Warning team. Collaborated with the team to invent KPIs for the org. This team optimizes detection strategies using existing signals and provides feedback highlighting detection gaps.
• Removed data scarcity by establishing best practices for leveraging moderation logs for training and evaluation datasets.
• Established modeling best practices. Transformers for text classification by pre-training on Tinder’s unique data; Text pre-processing pipelines to handle adversarial attacks; Metadata embeddings to improve multi-lingual performance; ConvNeXT family of models for effective and efficient image classifiers.
• Enabled real-time model inference. TFLite and Quantization for CPU-based models; TensorRT and Triton for GPU-based models; Collaborated with Infra team to establish requirements for the K8-based model endpoint framework.

Technical lead for the query understanding science team.

• Led the Research & Development of the first query -> shopping intent Recurrent Neural Network model. This model became the foundation for multiple shopping experiences that required mapping of queries to shopping intents.
• Established best practices for training Deep Learning models and their deployment for the team. Built GPU-based training and batch-inference infrastructure over AWS Batch.
• Collaborated closely with the infrastructure team to establish requirements for the inference framework. Made sure the team invested in a Python-based framework to make model deployments painless.

Founding engineer on the X-Ray team. X-Ray identifies a book’s topics, characters, images, and essential passages.

• Developed algorithms to identify characters and all their mentions in a book. For example, the algorithm will ensure that Mr. Potter and Harry Potter are associated with the same character Harry Potter.
• Led the Book Maps teams and developed algorithms to identify images and critical passages in a book.
• Developed testing at scale methodologies to go code complete months before the deadline. Concretely, each new algorithm change was tested against thousands of books, and the most common bugs were sorted by frequency and prioritized.
• Generated n-gram count statistics across the whole Kindle catalog using Map-Reduce and Dynamo DB. This data is used for topic modeling.
• Developed the Kindle N-gram Corpus. N-gram frequency table using Map-Reduce and DynamodDB across all Kindle books circa 2012. Essential topics in books are identified using this dataset.

• Text Classifiers for Small Scale Relevance
• Digital Content excerpt Identification
• Identifying entities in a digital work
• Identifying topics in a digital work
• Visual representation of supplemental information for a digital work • Providing supplemental information for a digital work
• Providing supplemental information for a digital work
• Presenting content in multiple languages
• Providing supplemental information for a digital work in a user interface
• Navigating supplemental information for a digital work
• Display screen having a graphical user interface for providing supplemental information of a digital work
• Layout-aware text extraction from full-text PDF of scientific articles