ML and engineering @Affirm. Interested in advancing fintech, AI in production, and Cricket! (views my own)
SF Bay Area
Joined August 2009
- Tweets1.7K
- Following268
- Followers596
- Likes3.7K
Pinned Tweet
1/ Super excited to share what we’ve been building @Affirm: a transformer-based underwriting model. A 🧵:
My bet is using other data belonging to consumers or limited data from bureaus, it will be possible to simulate how a mature credit profile will look like in the future.
Nitesh Kumar retweeted
this is like me and haircut
- i tried expensive barber (whisper flow)
- i tried to cut myself (holdtotalk.ai/)
- then i realized supercuts is all i need (Google AI Edge Eloquent)
😋
Six chefs. Six menus. A celebration of India’s culinary heritage this Diwali.
The Consulate General of India, San Francisco is partnering with six Indian restaurants in San Francisco for Indian Restaurant Week, November 1–14, 2026, highlighting the diversity of India’s regional cuisines and culinary traditions.
Discover the menus: indianrestaurantweeksf.com
@MEAIndia @IndianEmbassyUS @diaspora_india @IndianDiplomacy
A lot of engineering is aligning architecture to the customer problem beyond just internal convenience.
On-demand aggregation of user histories made checkout slower the more someone used Affirm. The team precomputed those features instead and kept the expensive path rare.
Latency reductions (precomputed vs on-demand):
We moved expensive feature computation out of the checkout path, then had to solve the harder question: how do you know a precomputed value is still right?
We wrote about the design, the tradeoffs, and what we learned: medium.com/affirmengineering…
not the kind of agent you'd expect in an OpenAI post
We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI, and we’re grateful to the agents and prosecutors who moved quickly to make an arrest. The FBI’s investigation is ongoing and we will continue to support their work.
At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data. The employee was terminated.
We take protecting our technology, data, and people seriously, and as a matter of national importance. Our security program is designed to protect against a wide range of threat actors, from insider risks to cybercriminals to nation states, and this case is an example of how AI can help address these sophisticated security threats.
AI needs to create a whole swath of new math problems that current AI cannot solve.
We have exhausted centuries worth of open math problems for PRs and benchmarks.
There's a lot about Muse being cute/ avatars/ personality (I call mine Chotu) and all that makes people trust it..
There's some truth in that, but the thing that hits me is the browser that enables you to see what its doing and invites takeing control -- thats trust building!
In the future, tabular foundation models will become nearly as ubiquitous as LLMs are today.
We hope this work we bring to the community enables modelers to get comfortable with adopting these at scale, and pull that future up.
Our tutorial on Tabular Foundation models was accepted to the NeurIPS 2026 Education Track! Give it a try here to see how a pretrained model uses a table as context: affirm.github.io/tabular-fou…
Really great stuff coming out of our ML and ZT1 teams at @Affirm
Explainability. You still need to explain why an application was declined. Stacking a transformer with XGBoost makes that harder. We developed a proprietary algorithm to do ensure we were able to do this with the full rigor! We ensured latency remained low, in miliseconds.
A lot of work from the team went into the modeling, data and production systems behind this. More on what we built and learned: Full post tech.affirm.com/learning-ric… (End of 🧵)
