As the first ML Engineer, I co-founded the modern machine learning efforts at Coinbase, acting as the team's tech lead for the last three years over a 40+ team. I had a role in building, designing, or advising every ML model running in production currently at Coinbase, as well as the technical and strategic direction. Highlights:

* Landed double-digit increases in site-wide net revenue and engagement across the Feed, Growth & Notifications, and user spend limits; strategic independence of 3rd-party vendors and a massive reduction in fraud for the ATO & Risk domain; and improving support agent response time.

* Created Auto-ML-like lego-blocks around tabular data, sequence & text data (using Transformers - the T in ChatGPT - from scratch & deep learning), and support for blending these models. Machine learning engineers and data scientists across the company use these lego-blocks to deliver impact.

* Set long-term technical direction for the team, owned large parts of the interview process, published a blog post around our stack, contributed to a KDD presentation, and ran a knowledge-sharing group and boot camp.

* Proposed a Responsible AI framework around model cards inspired by academia, which is used by new models.

* Working on the first ML + Crypto use cases.

Worked on applying NLP on Slack chat logs - expert finding, topic modeling, text classification. Improved search quality. Later on, built full-stack features such as web search and external note sharing.

Machine learning

===============

Expert finding:

• Tf-idf based model for recommending experts based on a query string

Topic modeling:

• Pipeline for cleaning text data, parsing out stop-words and first names, keeping only nouns

• Langid-based classifier for keeping only english-speaking teams

• Clustered the cleaned-up text messages using LDA in 50 clusters

* Visualized results using pyLDAvis

Text classification:

• Collected a labeled dataset for classifying messages into design/non-design

• Built additional filtering step for design-messages using a vocabulary from design-related Wikipedia pages

• Used fastText as an initial baseline (both with pretrained Glove word vectors and with training the word vectors). Got 0.83 f1-score

• Reproduced colleagues Bidirectional LSTM classifier using Keras. Got 0.86 f1-score

Technologies used: jupyter, pandas, fasttext (text classification), gensim (lda), keras (bidirectional LSTM)

Search quality

============

• Collected gold dataset of queries and expected positives for those queries

• Built search quality eval tool

• Iterated on the ElasticSearch model

• Improvement top3 recall from an initial 30% to 90% for that gold dataset

* Had follow-up call with an ElasticSearch outside expert consultant

Full-stack

========

• Web-search

• External note sharing (via public link)

• Invite flow

• Note Reactions (similar to the Facebook ones)

Technologies used: Python, React, CSS, HTML

* In charge of building the big data system in charge of making sense of terabytes of diverse information

* Worked on building a one box type of search system for the various data collected by BitDefender, using Cassandra, Spark and Django.

* Aim of the system was two-fold: easily search through terabytes of data and automatically generate forensic leads for the analysts to investigate.

* TA for Algorithms & Data Structures, Complexity Theory and Randomized Algorithms

* Held weekly TA hours for explaining difficult course material, at both undergrad and grad level

* Responsibilities also included grading homework