Works

HPC Cluster Simulator

A drag-and-drop, gamified training environment for building and operating GPU clusters. Trainees design a cluster, run workloads against it, break it, and learn why it broke.

Built around a custom conservative parallel-discrete-event core with seven coupled in-process engines — workload, scheduler, compute, fabric, storage, thermal and failure. Upstream simulators (Slurm, Astra-sim, Batsim/SimGrid) are used as offline validation oracles rather than live dependencies, so the runtime stays fast and self-contained. The fabric model is calibrated against Astra-sim; a 15-concept pedagogy catalog annotates every scheduling decision with what happened, why, and how it maps to real machines.

• CoreConservative PDES, 7 coupled engines
• SchedulingFCFS, EASY-backfill, fairshare, topology-aware
• FailuresMTBF Poisson, SDC, slow NIC, BMC hang, CRAC loss
• Front endReact Flow canvas + react-three-fiber rack theatre

The Immortal Daemon

An RL environment where the agent must keep a realistic cloud application alive while chaos batters it — and every reward is machine-verifiable, not judged by a model.

The application is an e-commerce checkout service running a saga against an AWS-compatible backend. A fault injector escalates through process kills, resource pressure, backend outages, network partitions and clock skew; the agent restarts, scales, drains, degrades and fails over to keep orders flowing. After each episode an invariant oracle reconstructs ground truth and asserts what a checkout system may never violate: no oversell, no double-charge, no lost orders, no money without goods, full fulfilment recovery. An agent that serves fast garbage fails the oracle.

• SubstrateCubeSandbox microVMs on GB10
• Fork cost~92 ms boot · ~230 ms snapshot · 16 clones in ~470 ms
• Cloudfakecloud — 19 MB Rust AWS emulator, ~47 MB RAM
• TrainingGRPO with group-relative advantages over identical forks

legal-financial-modernbert-150m

A 150M-parameter ModernBERT encoder trained from random init — no pretrained checkpoint, no distillation — on public legal and financial documents, then turned into a Matryoshka embedding model whose vectors truncate without retraining.

124 GB of deduplicated corpus becomes 25.79B tokens, 80% legal and financial against 20% FineWeb-Edu, under a 50,368-token vocabulary carrying 156 domain tokens. Stage-1 pretraining runs 271,000 steps to 71B tokens — 475 tokens per parameter — at roughly 28,400 tok/s in bf16 with sparse prediction and full torch.compile. The dedup pass is a Go program doing 604 MB/s, 24× the Python reference. All of it on a single desk-side GB10.

• Params149.7M, ModernBERT-base geometry, random init
• Corpus124 GB deduplicated → 25.79B tokens
• Schedule271k steps → 71B tokens → 2.76 epochs
• Hubsakshamio/legal-financial-modernbert-150m

Stage-1 pretraining in progress. Published checkpoints are early and no quality claims are made yet.

Populace

A world-scale persona simulation: grounded personas on a real globe living daily routines and reacting to events, with a language model authoring what each kind of person makes of the news and a Go engine doing the per-capita work.

One million personas, 7.5 million social ties and 450 archetypes tick in about 13 ms. On top of the social graph sits a media layer — ranked feeds as a structurally different contagion channel from peer ties. Amplification is modelled as engagement raised to 1/(1+A), which over-samples the tail: at A=3.4 a story circulating among 0.1% of users takes 21% of feed slots. The result is not "everything spreads" but a critical mass that moves — at 1.2% seeding, 4.12% of the population without platforms and 99.91% with them.

• Scale1M personas · 7.5M ties · ~13 ms/tick
• ViewsGlobe, morphing equirectangular map, group chat
• Model pathRailway → Cloudflare Tunnel → SGLang on the Spark
• Livepopulace-production.up.railway.app