atle.dev
atle.dev / about

Andreas Atle

Email me

I build AI systems with the habits of scientific computing: bounded behaviour, verifiable states, and claims backed by measurement.

My background is numerical. I have a Ph.D. in applied mathematics from Stockholm University and an M.Sc. in scientific computing from KTH. After a postdoc on finite-difference and eikonal solvers, I spent eleven years at Total E&P in Houston building high-performance seismic imaging and inversion algorithms with MPI and CUDA. I've been writing code since the late 1980s.

What I work on now

Learned world models and planning. In 2026 I built a JEPA world model from scratch: trained on pixels, probed for what it actually knows, and used for planning with CEM inside an MPC loop. It solves dm_control's ball-in-cup 97% of the time from random starts. I'm now testing whether hyperbolic latent spaces represent branching futures better than Euclidean ones. On a grid maze they clearly do at low dimension, and the result reverses at high dimension.

Reliable document AI. Multi-engine OCR for scanned title documents: LLM vision for prose and layout, deterministic OCR for numbers, reconciliation by vote, and domain verifiers for what voting can't catch.

Governed agents. Several years of LLM agent systems with typed state, explicit state machines, bounded retries and auditable traces, including production pipelines that processed 1,300+ legal documents.

How I work

  • Measure, don't assume. Build the diagnostic, run it, and report null results.
  • Explicit over implicit. If a system does something, I want to point at the line that made it happen.
  • Ablate your fixes. When a stack of changes works, find out which ones carry the result.

I'm looking for work in the agent space, especially where world models, planning and engineering discipline meet. The best way to reach me is atle.andreas@gmail.com.