maksis knutins
hi, I'm a generalist research engineer from Latvia. I've worked across the software engineering spectrum from low-level traffic controller integrations to training a 70B LLM from scratch to building the full stack of SaaS platforms. I'm interested in complex technical problems which create positive externalities in the world.
projects
-
sonora
a medtech startup doing autonomous medical coding and reimbursement optimization. I was tech co-founder, and got us to MVP with 97% acc. in DRG group prediction and 92% in reimbursement forecasting across two successful trials. simulations showed $2.1M+ annually in lost revenue per hospital in our region of interest, but we wound down after wrestling with the glacial market and finding a scalable business model.
post-mortem (todo) -
imbue-70b
we trained an LLM from scratch! I led data infra and processing, explored active learning, and made synthetic and scavenged code evals.
-
on the stepwise nature of SSL
a collab with Jamie looking at some weird things we saw training joint-embedding SSL models. I was the empiricist and built the experiment harness + baselines, ran experiments, matrix crunching and helped write the paper (icml '23).
-
CARBS
Abe made a cost-aware hyper-parameter optimizer. the paper didn't make NeurIPS, but we got a ton of mileage of it internally at Imbue and eventually open sourced it. I worked on benchmarking and implemented baselines (asha, bohb, blendsearch etc).
-
avalon
reliving the childhood dream of making a 3d video game, but for RL agents. I designed the procgen engine, tasks and gameplay for interior levels, as well as bits of the VR integration (yes, you could play it in VR).
-
smart junctions
we trained an rl traffic control algo on a simulation of the city and then deployed it to multiple junctions in Manchester, beating SCOOT (incumbent) by a good margin. I worked on our training framework, drove RL research - lit review, reward sculpting, environment engineering, simulation fidelity, etc. also helped with deployment + monitoring infra and tooling.
-
*city
at vivacity, all our repos were nouns that ended in -city. I led the analysis team, driving the data dashboard (audacity), a self-service app for processing pre-recorded videos (felicity), an in-house labeling tool (perspicacity) and many others.
-
manta equipment
an inventory management SaaS for universities with multiple clients in the US. it was my first hustle at building and growing something from scratch with my co-founder. still chucking away in the background!
-
cityseek
uni research project building a workflow and tooling to generate daylight availability models for cities from Google Street View data. I lacked the chops and time to throw deep learning at it, but nowadays it should be pretty easy!
-
mini transport tycoon
a quaint uni project in OCaml, combining two city games - transport tycoon (childhood fave) and minimetro.
contact
hi [at] (this domain)