Work

Hi there 👋 My name is Taha. Here I write about work, plus some broader stuff at the end. If we haven’t met yet, I hope this gives you a quick proxy for my skills and experience.

Current Terminal Goal:

I want to help humanity build AGI. I think we can build one. It’s super worth it to build one. We can all chill and be unique humans afterwards. That’s the world I want to live in. And not a world bounded by fear, shame, guilt and lack of agency as a recurring pattern of human experience.

People grow up playing games with each other and see others defect. It makes the brain overfit on those data points to believe that “kill consume multiply conquer” is an inherent quality of the territory – because that’s the loss function of evolution or something along those lines. It’s easy to be stuck in the bounds of a model of reality like that and really suffer. And totally miss out on the possibilities that emerge and manifest once you try a bit harder.

Things like software, language, thrust of stories from mythology, and structulism is creating this awesome playground for us to have a blast by being playful and explorative. It’s opening up a safe space to play, and robust patterns of meaning in the nebulosity of emptiness. Isn’t that all worth working past flincing away, and walking into the unknown?

< Focus />

I build data products — for the most part as an engineer, doing napkin math, designing large pipelines with lots of components, and coding robust and scalable things. But I wear my PM hat too here and there, which has helped with delivering useful things, bending risk, and asking for help.

Sabbatical ⚓️ _{^{(2023-09 to now)}}

(I am open to opportunities for ml related contract/full-time jobs at seed to growth stage startups, but I’m a bit slower to respond at the moment if you’re reaching out)

I’m taking a long break. Planning for up to 6 months, but I will stay open to start sprinting again if I feel motivated by a project. So far I’ve been broadly exploring lots of interests. Not making any bets yet. I will write about my Explore/Exploit plans soon. The general theme is that OODA loop is more important than ever at the moment.

• Been reading a lot! fermi questions, epistomology, information theory, minimal desing, open source developement, scaling infrastructure.
  
• I’m meeting with ai alignment friends every week, exploring alignment problems.
  
• Also working on a Mech Interp research project inspired by recent Neel Nanda’s work.
  
• I went on dates with some wonderful people, and had deep conversations.
• I did probably my longest interview ever at Goldman Sachs (~ 7 hours lol) for a gig to scale their LLM inference on their large stack of H100s. I had a great conversation with their ML lead, doing napkin math going up and down the stack. I never felt as excited about a gig before!
• My friend Mike (mikador64.com) runs sys admin things for a major university. I’m building him an ai terminal to automate most of his sys admin tasks away. Moving very fast on this so far.
• Playing code golf with programmer friends (competing to write a program in fewest characters possible).
• Kayaked very long distances at Lake Ontario (I’ve worked up to +8 hours so far). Also there’s a swimming spot at the end of my route on back side of Ward’s island.

MLOps Engineer @ Flinks _{^{(2022-10 to 2023-08)}}

[Flinks raised a $103M series B]

Continue running our data at Flinks ( Data warehouse and most data pipelines ), as well as our MLOps now, supporting the DS team flinks.com/attributes – Two recent tangible projects:

1) Migration of our ML pipelines from AWS to GCP (SageMaker to VertexAI)

2) Led development of our largest ML pipeline, ingesting billions of banking transactions close-to-realtime, generating predictions using 6 different ML models in a sequence, and reliable delivery of the aggregated results to client facing dashboards (team effort of 9 engineers and data scientists )

Tools:

General tools: Python, Bash, Docker, DBT, Airflow, Flask
GCP: VertexAI(Kubeflow), BigQuery, Dataflow (Apache Beam), Cloud Functions, Cloud Composer (Airflow), PubSub, Cloud Storage, MemoryStore (Redis), FireStore
AWS: SageMaker, Lambda Functions, S3, Redis, API Gateway, CloudWatch 
CI/CD: Azure DevOps, Terraform

Data Engineer @ Flinks _{^{(2022-05 to 2022-10)}}

Built our new data warehouse (on BigQuery) and ETLs, currently powering most of data at Flinks:

• Built our entire DBT project — A clean breakdown of +100 models serving different teams and pieces of product, in multiple dev/prod environments, orchestrated with Airflow, and unit-tested in CI/CD, and data quality tests in great_expectations

• Replaced our slow and expensive API data ETLs with a streaming pipeline, ingesting data across many sources — Eliminated 24 hours delay for API data to reach the data warehouse, and enabled us to matrialize DBT models close to real-time

• Wrote the backend of end-user event tracking API in Golang, streaming to BigQuery using GCP Pub/Sub and Apache Beam (deployed on GCP Dataflow) — Improving capture rate of end-user sessions from avg of ~ 30% to 90%+ across all customers, processing order of magnitude of ~5M events/day

• Wrote our PII obfuscation layer — Made it possible to open up access to billions of rows of sensitive banking data, to various teams internally

• Helped replace thousands of lines of the complex code-base for two major parts of our internal products (billing and banking connection load balancing) with simple SQL models

Tools:

Data tools: dbt, great_expectations, lightdash
General tools: Python, Golang, Bash, SQL, Docker, Airflow, Flask
GCP: BigQuery, Dataflow (Apache Beam), Cloud Storage, PubSub, Cloud Functions, Cloud Run
CI/CD: Azure DevOps, Terraform

Product Data Scientist @ Flinks _{^{(2021-02 to 2022-04)}}

Built many of our product analytics models and dashboards (SQL, Pandas, PowerBI, Streamlit) — These are frequently being used by various teams today.

Trained statistical models around fluctuations in quality of US/Canada banking connections — This lead to useful insights around major drops in conversion rates for some banks, impacting order of magnitude of ~ %10 of ARR.

Built a metrics dictionary for the company, with the help of our PMs and engineering leads (really grinding work, but rewarding) — I later used this for data modeling of our new data warehouse.

I then got support from many stakeholders for designing and building a new centralized data warehouse on BigQuery, streaming ETLs, and DBT models downstream — This is powering entire data at Flinks today.

Tools:

General tools: Python, Bash, SQL, Docker, Streamlit, pandas, numpy
GCP: BigQuery, PubSub, Dataflow (Apache Beam), GCS
CI/CD: Azure DevOps, Terraform

Software Engineer @ AI Research Lab _{^{(2020-09 to 2021-03)}}

Worked on a ML storytelling engine. Wrote code for preparing large datasets, training and fine-tuning various models, as well as serving them:

• Image captioning with PyTorch
• Fine-tuning GPT2 models on GCP GPUs
• Generating melodies based on text sentiment
• Generating images using Open AI’s DALL-E
• Wrote a REST API for serving these models with Flask, and a React Native app

Tools:
Python, Bash, Docker, PyTorch, GCP, Flask, React

Full-Stack Dev @ Zaxada _{^{(2019-05 to 2019-09)}}

Wrote the MVP for an Austrialian seed stage startup. An e2e E-commerce integration platform to connect indvidual stores (Shopify, Woocommerce, Magento) to Southeast Asian marketplace (jd.com, zilingo.com). This included componenets for Orders and shipping management, Inventory management, Product management, and Pricing management. Backend in Django and Frontend using boilerplate react.js code.

Tools: Python, Bash, Docker, Django, Gunicorn, NGINX, PostgreSQL, Celery Beat (with Redis), React.js

Product Manager @ ArochoTech _{^{(2018-04 to 2018-08)}}

Helped build a real-time prediction app for world cup. I led product, working with 4 devs on the team. Few tangible things I did:

• Designed the app from scratch around a single core feature, maintained product backlog and priorities
• Brought in one of the most well-known gamification designers in Iran, and learned a lot about UX desgin from him
• Handled the legal contracts for our new mobile payment methods (this was very tricky from a legal standpoint)
• Helped marketing with their planning for tracking user acquisition and retention

Data Scientist @ Banimode.com _{^{(2017-11 to 2018-04)}}

Biggest clothing e-commerce store in Iran in terms of market share. I helped marketing team efficiently grow traffic very quickly (~3x in 6 months)

• Built models for optimizing cost of display ads, paid search, socials
• Helped keeping CAC low while the team scaled up ad spend
• Designed and implemented a marketing automation system - enabled the team to run and scale many personalized retention campaigns
• Wrote queries for extracting data from prod dbs of our e-commerce store, to measure various marketing metrics (Python, Bash, SQL, Pandas)

Co-founder and Product @ Poloto _{^{(2016-08 to 2017-09)}}

Co-founded an ecommerce platform in Iran (similar to Shopify), and helped grow it to ~50 paying customers and grew the team to 12 engineers. I lead the product, working with some wonderful people, pivoting to find product market fit. And made many PM mistakes as my first experience working with devs, so I learned a lot. We did show traction, but eventually couldn’t make our subscription model scale.

Externally, I helped raise an angel round, and gave investors a good idea of how we chose to deploy our runway. Also paulgraham.com/love.html

.:~:..:*~*:..:~:..:*~*:..:~:..:*~*:..:~:..:*~*:..:*~*:..:~:..:*~*:..:~:..:*~*:..:~:..:*~*:..:~:..:~:..:*~*:..

Eduction: autodidacticism for most part.

I grew up in Iran, and went to a top engineering school in 2016 to study embedded systems ( ranked 405 among ~ 180k participants of Iran’s mathematics university entrance exam competition) - But then skipped most of the classes, occasionaly sat at some grad courses, played with micro-controllers, taught embedded programming to undergrads, fixed a very expensive robotic arm, made some money freelancing, and eventually dropped out to build a business.

I then moved to Toronto, and took some CS, math and stats courses. Lots of good Data Structures/Algorithms practice, grad level stats courses, some heavy math ML courses, wrote a DNS resolver, and implemented OS components in C. Also played lots of poker with friends.

.:~:..:*~*:..:~:..:*~*:..:~:..:*~*:..:~:..:*~*:..:*~*:..:~:..:*~*:..:~:..:*~*:..:~:..:*~*:..:~:..:~:..:*~*:..

Broader Life and Play

I currently live in Toronto, and I love the city. Prior to that I grew up in Iran. I consider myself a rationalist and buddhist. I enjoy playing poker with friends, listening to 90’s rap, and reading sometimes (recently mostly on topics like epistemology, information theory, philosophy, rationality, quantum physics, game theory, etc.). Recently I’ve been developing an interest in building quantitative models for systems with uncertainty in them (read financial markets). Lately I’ve been spending my time outside of work on quantitative trading research, reading research and implementing research tooling.

Intuition

Here are some variables I sometimes think about, working on projects. They’re not principles, but just loosely held opinions/stances (hopefully):

• It’s important to do things fast
  • slow is fake
  • doing things faster leaves no time for things that don’t really matter
  • people burnout when they don’t feel the momentum
• Energy matters
  • does the thing you’re working on draw energy or add energy
  • make changes if things don’t feel right (sometimes this is hard, so stay patient as long as it’s not learned helplessness / lack of agency)
  • remember that it is possible to be in flow for long streches of time everyday
  • happiness and productivity do correlate more often than not
• What not to build
  • judgment matters a lot when you have leverage
  • code costs you. rarely ever it is well justified to write garbage. be mindful of this.
  • get into the habit of picking up slack all the time. it helps with all kinds of internal resistance
  • don’t build something people don’t need. let’s get real or let’s not play
• Smaller teams are better
  • less meetings, more opinionated direction and momentum
  • less resistance towards downside of trying things out. faster decision making
  • faster feedback loops
• Mistakes are not bad
  • as long as you avoid risk of ruin. keep enough chips left to stay in the game and bet again
  • optimize for expected value of sample information
  • often experimentation works better than thinking about things for too long
  • your goal is gaining in EV. learn to work with risk of downside, when there’s appealing upside
  • when you catch one of these mistakes be okay with climbing down and correcting for it

Zen of Python

>>> import this

The Zen of Python, by Tim Peters

Beautiful is better than ugly.
Explicit is better than implicit.
Simple is better than complex.
Complex is better than complicated.
Flat is better than nested.
Sparse is better than dense.
Readability counts.
Special cases aren't special enough to break the rules.
Although practicality beats purity.
Errors should never pass silently.
Unless explicitly silenced.
In the face of ambiguity, refuse the temptation to guess.
There should be one-- and preferably only one --obvious way to do it.
Although that way may not be obvious at first unless you're Dutch.
Now is better than never.
Although never is often better than *right* now.
If the implementation is hard to explain, it's a bad idea.
If the implementation is easy to explain, it may be a good idea.
Namespaces are one honking great idea -- let's do more of those!