About

I'm Tsolmondorj Natsagdorj, a systems and research engineer. I work across reliable AI, high-performance networking, scientific ML, autonomous systems, and security-critical infrastructure.

The through-line is correctness under hidden failure modes: benchmarks that leak their labels, mocks that agree with the bug they are testing, uncertainty estimates that do not track error, protocol invariants that prove less than their diagrams imply, and ambitious hardware that needs physical gates before performance claims.

Formation explains how protocol invariants, evaluation validity, calibration, hardware boundaries, and physical autonomy became one engineering trajectory rather than five unrelated interests.

Now

I build Suwappu and related systems, work on BRIDGE-bench and AI-evaluation methodology, maintain scientific-ML experiments around materials discovery, and prototype Aiur, an autonomous airborne-carrier research program. I also contribute correctness fixes and infrastructure upstream when a problem belongs in the shared library rather than only my project.

Selected upstream

• MatGL #801merged — reusable MC-dropout uncertainty primitive with explicit supported-model boundaries.
• MatGL #809merged — SoftExponential autograd and numerical-domain correctness.
• MLX-LM #1372merged — XTC sampling default correctness across sampler, CLI, and server.
• Alloy #1105merged — EIP-712 self-referential type canonicalization while preserving strict runtime-resolution limits.
• uutils/coreutils #12327merged — GNU-compatible `date` timezone re-zoning.
• uutils/parse_datetime3 merged — AM/PM composition, negative fractional epoch floor semantics, and `UT` timezone compatibility.

90-second upstream evidence view →

Research practice

I separate observed, derived, estimated, and target claims. A later validity failure is allowed to weaken an earlier result. That is why the materials and BRIDGE-bench articles carry visible revisions instead of preserving the flattering interpretation.

public revision ledger · research programs · engineering thesis

Before

I co-founded Blockchain at Mason in 2016–2017, a student-run, faculty-supported organization that grew to 90+ members. I studied Economics & Computer Science at George Mason University and left in 2017 to work full-time.

Elsewhere

github · x · linkedin · orcid · suwappu.bot