I lead the engineering organization for HubSpot’s AI product line. We use AI to help millions of businesses grow better.

I run my one-man quantitive investment firm, focusing on taking advantage of what I call “narrative discontinuity”.

I now lead an initiative called MVAI (minimal viable AI) inside Meta. We take a startup mindset to build a brand new AI stack from scratch. Previously I led two large cross functional efforts to move FB's recommendation models (models behind FB Reels) to GPU training and inference. Both increased Reel's top line metrics by more than 10%.

I manage Feed, Search, ML Platform and Applied ML (modeling) teams at Reddit. Our teams are responsible for core product surfaces (Feed+Search) at Reddit. The teams are also responsible for building the ML platform (feature store, training pipeline, serving infrastructure).My contributions:* put together Reddit’s first ML vision & strategy doc, presented to the CEO and put it in action;* led the team to deliver the first ML driven home/popular feed for Reddit;* led the team to deliver the first (Spark) and second generation (TensorFlow) ML platform for Reddit;* led the development of latent models for user, subreddits and contents, powering experiences such as subreddit recommendation and push notification;* collaborations with other teams to proliferate ML across the organization, for Ads, Growth etc.

Basically the idea and algorithm behind Github CoPilot / OpenAI Codex, but five years ahead of its time. Applied AI (LSTM and large language models) to solve developer productivity. Trained a bunch of neural networks to understand and generate programming code, used large-scale language models and extensive compiler technology to do paragraph-level code autocompletion, with the ultimate vision of "computer, write code that does X". The tech stack, for curious eyes: Google Cloud Platform (Pubsub/Dataflow/Spanner/BigQuery), Tensorflow/Keras/Edwardlib (yes, compiler + Bayesian approaches combined with deep neural network). I trained a few DNNs on a large portion of the open source code ever written by human race, hosted on our own source-control backend (basically developed a Git backend on Google Cloud, optimized for ML tasks).The real IntelliJ plugin + demo: https://plugins.jetbrains.com/plugin/9203-ai-predictive-codingThe work predates Transformer/GPT, and language models of that era were far smaller than modern LLMs (I trained models with 400M parameters or so). Alas, it was great to build around the vision that AI can be used to generate code.

* Grew and managed the sleep/wellness engineering team (essentially the Fitbit ML team) 3x, from 5 engineers to a total of 16.* Led the development of two generations of machine learning platforms at Fitbit, from design to production.* Architected systems to split a deep neural net into wearable hardware (DSP) and cloud, to save power and communication while retain model accuracy. * The products I led and architected, Sleep Insights and Sleep Stages, launched in March 2017. Sleep tracking remains one of the most popular Fitbit features today.

Fitbit's SRE team was formed after an embarrassing system meltdown during Christmas time, 2014. I volunteered to join and becoming a founding members of the SRE team after the crisis. I was the tech lead directly under a VP.From scratch, we defined our mission, wrote the team charter, and hired more engineers. A few highlights in my tenure as an SRE lead.* identified and fixed a boat of scalability problems around memcached, MySQL scaling & load balancing, significantly advanced tooling, logging and monitoring around production systems. As a result, Christmas time 2015, everything was peachy.* led the team through company IPO. There were surprisingly a ton of work for that (e.g. compliance, security).* transformed a chaotic release process to a regular, reliable release schedule, by developing tools, optimizing process, and coordination with a ton of stakeholders.

Together with my team, we manage a data pipeline that processes all Google's display ads data -- that is about 20% of all display ads of the whole Internet.I led a team of four engineers (+PM) in developing a product later became known as Data Driven Attribution (http://doubleclickadvertisers.blogspot.com/2016/02/enable-better-decisions-with-data.html). The system utilizes vast amount of data to distribute "credits" to ads channels. The pipeline was built on statistical machine learning and contrafactual reasoning. In my last year at Google, I also did stint for the self-driven car team (remote). My job was to "drive" the self-driving car in Google data centers (i.e. simulation) to ensure software quality. I definitely became a better human driver after "driving" these self-driving cars in the Cloud.