Kartik Joshi
Dallas, TXI’m interested in ML research and AI engineering.
The research half is model training, mostly post-training. Using RL to align a model so it gives better answers. That is what the CAIR Lab work is, and what most of my own RL projects are about.
The engineering half is building the systems that run models in production. That is close to what I did at Goldman this summer, and I build open source in the same direction: agent control planes, evaluation harnesses, and tooling for the work.
instagram and x are where the daily stuff goes, mostly whatever I am building that week
I grew up in the Bay Area, where roughly everyone’s parents worked at a company you have heard of. Being a kid there and not being into tech was basically a personality choice, and I did not make it. I started writing code in high school mostly to find out whether the hype was real. It was.
College started with web development and a run of projects that had no theme beyond me wanting them to exist. A decorator for figuring out a dorm room. A couple of dating apps aimed at campus life. None of it was strategic. I just liked that you could think of something on a Tuesday and have people using it by Friday.
The summer after freshman year I got into research and it took over. An NSF REU on multi-agent security, which became a paper. It was the first time I worked on a question nobody already had the answer to, and I have not really wanted to go back since.
What stuck from it was smaller than the paper. A model writes an equally convincing justification for a wrong answer as for a right one. So the answer was never the interesting part. Whatever sits downstream and can tell you the answer is wrong, that is the interesting part.
From there it went into deep learning and reinforcement learning, and then into the engineering side of making agent systems that hold up under real use. At Goldman I built agents whose whole job was attacking findings other agents produced. Novus came out of the same instinct, that a team running coding agents needs evidence rather than vibes. RL from verifier reward turned out to be the same idea moved inside the training loop.
Right now I am at the CAIR Lab on continual learning for video diffusion, still building Novus, and still pushing on RL post-training. More fired up about all of it than I have ever been.
Basketball
Warriors, and past that, Steph. I have watched him do things that made no sense at the time and then became the way everyone plays.
Movies
I got into film properly only recently and I am making up for lost time. Right now I am working through Nolan’s filmography front to back.
Nature photography and travel
I shoot a lot outdoors. Half Dome from Glacier Point, the red rock in Sedona, the reservoir at Pinnacles, the Northern California coast. Japan is next on the list.
Food
I photograph what I eat, which sounds insufferable right up until you see the pictures. Mostly I am working through the Dallas scene one place at a time.
Twitch
An unreasonable amount of it. I think it is a great way to consume content.
Pokémon cards
Started as nostalgia, became a spreadsheet. The whole collection is public on Collectr as @sn3hyx.