Deterministic cores for nondeterministic systems.
Tools that make concurrent, AI-assisted, and parallel systems replayable and measurable.
Developer tool that automates database race testing in Spring Boot CI
- Replay · saves the interleaving between sync-points and reruns it against real MySQL.
- Judge · SQL invariants decide pass or fail, so the verdict does not depend on timing.
- Prove · on one saved schedule, the vulnerable Spring Boot app fails and the fixed one passes, 20/20 each.
- Gate · one workflow step, published on GitHub Marketplace, fails the pull request with the schedule, trace, and violating rows attached.
Deterministic replay Go MySQL 8 / InnoDB Testcontainers GitHub Actions
Deterministic trade-case replay with measured selection of its column-mapping model
- Propose · a model maps each column of an unfamiliar trade file to a canonical field, or returns
REVIEW_REQUIRED. - Check · a closed contract and a dry run reject invalid proposals; a person approves the exact proposal, bound to its hash.
- Compute · the replay engine re-derives events from stored rows and tests a versioned pattern rule; each finding traces to its
rawRowHash. - Select · candidates run on a sealed held-out set beside a non-model baseline; six criteria fixed before the run decide whether any may propose mappings.
AI trust boundaries Model evaluation TypeScript Next.js
| Performance studies | What it measures |
|---|---|
| surveillance-frontier | Time-budgeted stress testing of market-surveillance detection policies on CPU and GPU. The incremental exact method recomputes only the windows a test event touches, cutting a trial from 1,180.55 ms to 3.36 ms (349×, one week of 1-second BTCUSDT bars, one thread), with 0 mismatches against full recomputation across 2,112 checks (C++ / OpenMP / CUDA) |
| drone-doa-gpu | Real-time acoustic direction finding for drones on an 8-mic array. On a Tesla T4, moving the whole pipeline to the GPU cuts 99.4 ms to 2.58 ms per 64-frame batch (Python / CUDA) |
Team projects · built with teams across embedded, backend, and AI
| Project | What it does |
|---|---|
| Ongi Device | Connected medication-dispenser firmware, from caregiver schedule to confirmed intake. The device runs the server's schedule on synced time, opens the right slot, and reports whether the pill was actually taken (C / ESP-IDF) |
| Team-po / Server | Team matching and collaboration for developer side projects. Matching runs in a locked transaction so no one lands on two teams, and each team gets a group chat, GitHub integration, and an AI-generated development guide (Java / Spring Boot) |
| CarryCheck | Hybrid RAG agent for airline baggage rules, with deterministic decisions and guardrailed generation. On a 10-question set, BM25 and Qwen3 embeddings fused by RRF reach Recall@3 1.00, a harness that rejects changed statuses or unknown source IDs passes 10/10, and selective context cuts one request's tokens by 38.6% (Python / FuriosaAI) |
Foundations · coursework and labs in kernels, file systems, storage, compilers, and cloud platforms
| Project | What it involved |
|---|---|
| xv6-kernel-extensions | Kernel changes for saved file versions, observable physical memory, and weighted scheduling in xv6. Copy-on-write snapshots allocate 98.85% fewer extra blocks than a full copy (3 vs 262), and stride scheduling holds 1:2:4 CPU shares within 0.14 points, against 23.81 for round-robin (C) |
| linux-filesystem-tools | File organization tools from interactive copy to a background daemon, plus a read-only ext2 image inspector. Organizing 512 files / 128 MiB takes a median 0.122 s, with all 2,560 copies matching by SHA-256 (C) |
| storage-io-lab | Flash I/O, a hybrid flash translation layer, and fixed-slot record storage on a file-backed emulator. Reusing 250 deleted slots keeps a 1,000-record file at zero growth, and write amplification measures 1.000× to 3.988× by workload (C) |
| compiler | Eight steps from an expression calculator to a C-subset compiler. RD and LR parsing, scoped symbol tables, and type annotation lead to stack-machine assembly that the included interpreter runs (C) |
| cloud-native-labs | Concept notes and hands-on labs on how VMs, networking, and shared storage hold up a container platform. Covers Kubernetes workloads, Services and Ingress, persistent volumes, scheduling and rollouts, CI/CD, and AWS EKS (Kubernetes / AWS) |