resume · August 2026Tsolmondorj Natsagdorj
Systems & Research Engineer
I find where an abstraction stops matching reality, then build the test that exposes the gap. My work spans AI evaluation, high-performance networking, scientific ML, protocol correctness, and autonomy, with an emphasis on independent evidence, explicit limitations, and public correction when later validation weakens an earlier result.
Selected technical evidence
roce-preflight systems · networking
Real RDMA execution exposed four defects after 197 unit tests passed; the first attempted fix still failed on-device. Built a Soft-RoCE CI boundary that asserts no simulation fallback, exercises real verbs/sysfs state, and sends perftest traffic.
case study → · source ↗
BRIDGE-bench AI evaluations · security
Audited a real-exploit benchmark and found severe author-injected prompt leakage in 13/24 examples. Downgraded the historical model score to a contaminated upper bound and added sanitizer invariants plus a locked paired remeasurement design.
case study → · revision → · source ↗
active-materials-discovery scientific ML
On a static 18,928-structure perovskite screen, pretrained mean ranking reached DAF 5.15 at a 5% budget; the tested MC-dropout uncertainty signal had Spearman σ↔|error| = −0.47, and overweighting it pushed acquisition below random.
case study → · source ↗
Merged upstream — independent review
MatGL #801 ↗Reusable MC-dropout uncertainty primitive with explicit model-support boundaries.
MatGL #809 ↗Autograd-correct, numerically safe SoftExponential activation.
MLX-LM #1372 ↗Corrected pathological XTC sampling defaults across library, CLI, and server.
Alloy #1105 ↗Self-referential EIP-712 type canonicalization while preserving strict runtime recursion limits.
uutils ↗GNU-compatible date/time semantics across coreutils #12327 and parse_datetime #284/#285/#287.
Current engineering
Suwappu
Build distributed systems, cryptographic protocols, transaction infrastructure, and adversarial-reliability tooling.
suwappu.bot ↗
Aiur current experiment
Autonomy/hardware program reduced to a falsifiable recovery prototype. Current artifacts include design, CAD, simulation, controller logic, and acceptance gates; physical recovery performance remains explicitly unproven until telemetry exists.
source ↗
Formation
George Mason University — studied Economics & Computer Science; left in 2017 to work full-time. Blockchain at Mason — co-founded the student-run, faculty-supported organization in 2016–2017; grew to 90+ members. Full formation →
Selected writing
197 passing tests, four real-hardware bugs · A benchmark can measure its own metadata · Upstream is a different kind of test
Problems I want to work on
I’m most interested in teams where correctness has to survive contact with reality: AI evaluation and inference systems, low-level networking and distributed infrastructure, scientific ML, protocol/cryptographic systems, and autonomy where software eventually has to cross a physical boundary.
layerinfinite@gmail.com · 0xsoftboi.github.io · github.com/0xSoftBoi