Skip to content
View asp53826's full-sized avatar

Highlights

  • Pro

Block or report asp53826

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
asp53826/README.md
FAULT CHAMBER: a five-node Raft cluster is partitioned, elects a leader, repairs conflicting history, and emits a verification receipt

▶ ENTER THE LAB  ·   BREAK A RAFT CLUSTER  ·   INSPECT THE SOURCE  ·   RÉSUMÉ  ·   EMAIL

I build systems that can show where—and why—they fail.

The 30-second version

  • What I build: systems in C++ and Python, including consensus, storage engines, query planners, LLM inference and GPU compilers. Each one ships with a test suite, a named baseline, and the regime where it breaks.
  • Strongest evidence: raft-mvcc passes 598 assertions under seeded faults · vllm-lite reaches 2.9× throughput at 94% KV-cache use · track-fusion cuts localization error 47% in its published regime.
  • Now: UGA computer science, graduating December 2026 · industrial data engineering at MP Equipment · open to full-time systems and ML infrastructure roles.

Animated terminal: whoami, then make test in raft-mvcc passes 598 assertions, a FAULTLINE leader isolation produces two leaders that heal into one, and current status. A second pane shows a live lab monitor.

Now shipping: the latest human commit across the public portfolio

Enter the working lab

The live site asp53826.github.io/lab, filmed in a real browser by CI: an aurora field steered by the cursor behind the line 'I build systems, then I try to break them.'

↑ Not a mockup. A GitHub Action opens the live site in a real browser, steers it, and re-films this window whenever the site changes.

Inside the mechanism

FAULTLINE: isolating a Raft leader produces two leaders in different terms; the invariant monitor flags the commit-index split, then healing repairs it SIGNALROOM: measured replay of an IMM tracker on a maneuvering target and crossing tracks, with truth, measurements and estimates
FAULTLINE: the real C++17 Raft + MVCC engine in WebAssembly. Isolate the leader, watch two leaders appear in different terms, then heal and verify repair. SIGNALROOM: measured replays of the published tracking results, with truth, measurements and estimates kept separate, including the regime where the estimator loses.

The arcade

Arcade high-score screen built from measured results: 598 assertions, 138x warm read, 2.9x throughput, 47% less error, 68 tests. A pixel ship turns split-brain, stale-read and lost-write bugs into regression tests.

Short Gamma gameplay: hedging a short-volatility options book as delta swings, Greeks from the aad-greeks reverse-mode engine Order Book Royale gameplay: eight traders on one live order book on the lob-market-making matching engine
SHORT GAMMA ▶: you inherit a short-volatility book with ten days to the close. Every Greek on screen comes off the aad-greeks reverse-mode tape, and one number is wrong on purpose. ORDER BOOK ROYALE ▶: eight seats, one live order book, four minutes, last trader standing. It runs on the lob-market-making matching engine.
Evidence stack: claim, attack, oracle, boundary, and source-bound receipt

Execute the evidence

Open WITNESS, a live reproducibility chamber for eight receipt-bound engineering counterexamples

WITNESS verifies eight canonical receipts, runs each pinned C++/WebAssembly, TypeScript, or Python implementation inside the browser, compares the observed trace with its oracle, and emits a local run receipt. The local receipt is explicitly unattested; release provenance remains bound to COUNTEREXAMPLE.

Read the 23-page technical monograph · Inspect the GitHub edition · Download the Word edition

The publication formalizes WITNESS as an executable-evidence system through a threat model, eight capsule studies, eleven original diagrams, artifact-review readiness, explicit limitations, and a pinned identity register. It is independent technical documentation, not an awarded degree or peer-reviewed thesis.

Vidur: speculative decoding

Explore the results · Inspect the source · Read the modeling notes

An independent extension of Microsoft Vidur for proposal generation, target-model verification, accepted-token progress, and lookahead KV memory. The interactive page explores a 39-row simulator sweep where the verify-attention assumption can reverse whether speculation appears beneficial. These are predictions; real GPU validation is pending. The repository includes a matched-workload vLLM validation harness and 68 passing unit and fake-server tests; one additional end-to-end test requires a trained predictor cache.

Choose your route

Open Systems route Open ML Infrastructure route Open Autonomy route Open Quant route

Each route is a short, role-specific path through live systems, source, tests, measured evidence, and the regime where the result stops holding.

PROOFGRAPH is the interactive router across those routes: choose a hiring lens, then traverse one claim through its attack, oracle, measurement, boundary, exact source revision, and signed receipt.

Eight proof cartridges

COUNTEREXAMPLE public failure register Raft and MVCC consensus under attack
TensorForge browser tensor compiler SignalRoom autonomy mission control
EDGAR MCP bounded financial data vLLM-lite paged inference scheduler
FaultGraph causal incident workbench Portfolio Ops autonomous engineering control plane

Current flagship: COUNTEREXAMPLE

Twelve losing regimes from eight public engines live in one constrained failure register. Each capsule binds the mechanism, attack, oracle, limitation, exact source revision, downloadable receipt, and GitHub-signed stable release.

Operate the register · Inspect the source · Verify the release · Submit a bounded counterexample

Current release: FaultGraph

FaultGraph is a causal incident analysis workbench for operating on evidence rather than decorating a postmortem. It combines a FastAPI service, typed event contracts, checksum-addressed SQLite experiments, server-sent events, and a React/TypeScript forensic console.

surface released evidence
Model transparent dynamic linear influence model with documented assumptions and bounded do(source=0) counterfactuals
Workbench dependency graph, event timeline, ranked evidence, competing hypotheses, replay ledger, and keyboard command palette
Verification 11 backend + 4 frontend tests, strict type/lint gates, container build, CodeQL, and zero open security alerts at release
Supply chain signed v0.1.0 wheel, workbench archive, checksums, and GitHub attestations

The model is deliberately bounded: it ranks evidence and executes explicit interventions; it does not claim unrestricted causal discovery from arbitrary observational logs.

Inspect the workbench · Read the architecture · Audit the research protocol · Verify CI

Three reproducible passports

system mechanism measured proof reproduce
raft-mvcc Raft + serializable MVCC + linearizability checking 598 assertions across seeded faults; 11-tick five-node failover make test
edgar-mcp bounded SEC tools + cache validation + global pacing 32 tests; 138× warm 10-K read; SEC ceiling enforced make test
track-fusion IMM + JPDA + track scoring + OSPA 47% lower localization error in the published winning regime; failure sweep included python -m pytest -q
Clone and run all three
git clone https://github.com/asp53826/raft-mvcc && cd raft-mvcc && make test
git clone https://github.com/asp53826/edgar-mcp && cd edgar-mcp && make test
git clone https://github.com/asp53826/track-fusion && cd track-fusion && python -m pip install -e '.[dev]' && python -m pytest -q

The engineering contract

Build the mechanism. Attack the assumption. Measure against a named baseline. Publish the boundary. Bind the evidence to source.

The portfolio is organized as an engineering laboratory, not a gallery of screenshots. The interactive pages expose controls; the repositories expose implementation and tests; the evidence manifest records commands, units, and limitations.

instrument what it exposes
FAULTLINE a real C++17 Raft + MVCC engine in WebAssembly, including minority leaders, conflict repair, and passing or failing histories
KERNELARENA typed tensor IR, fusion, liveness-aware memory reuse, generated WGSL, and a browser-local oracle
SIGNALROOM truth, measurements, residuals, uncertainty, and the manoeuvre regime where an estimator loses
MARKETWIRE toxicity, quote age, inventory control, deterministic shocks, and a committed 20,000-step benchmark
WITNESS receipt hashing, pinned browser execution, deterministic oracles, tamper detection, and local run receipts
Benchmark Observatory exact commits, commands, environments, units, baselines, and limitations
Demo Cinema four captioned engineering films under one minute with direct routes into the running system
Open the complete source catalog

Evidence + reliability

faultgraph · counterexample · portfolio-ops · witness · proofgraph

Correctness + storage

raft-mvcc · dst-harness · hotstuff-bft · wal-recovery · lsm-tree · columnar-engine · query-planner · cdcl-sat

ML infrastructure

tensorforge-webgpu · vllm-lite · annlite · dist-train · feature-store · rag-eval · grammar-decode · ptq-budget · agent-harness · codebase-qa

Signal + autonomy

sdr-receiver · sar-focus · track-fusion · vio-nav

Financial data + markets

edgar-mcp · xbrl-normalize · lob-market-making · aad-greeks · backtest-honest

Inspect source-backed telemetry and delivery Portfolio telemetry generated from public repository source Source-backed development and verification toolchain

The daily workflow reads public source and counts test functions without a third-party stats card or page-load API call. Parametrized and looped suites execute more checks than the function count—for example, raft-mvcc alone executes 598 assertions.

Autonomous operations

Portfolio Ops maintains the public systems through eight auditable roles: repository verification, benchmark evidence, dependency care, documentation checks, demo monitoring, verified releases, achievement-state monitoring, and profile curation. Automation uses the GitHub Actions bot identity, retains raw evidence, and does not manufacture human contributions or interact with third-party repositories.

Patch dependency updates may auto-merge only after repository checks pass. Source changes, major upgrades, external contributions, benchmark claims, and account or legal decisions remain outside the autonomous boundary.

Project Primary language Latest release Latest completed workflow
faultgraph Python v0.1.0 Autonomous Engineering Lab: success
raft-mvcc C++ v1.0.0 Autonomous Engineering Lab: success
edgar-mcp Python v0.1.0 Autonomous Engineering Lab: success
track-fusion Python No published release Autonomous Engineering Lab: success
tensorforge-webgpu TypeScript No published release Autonomous Engineering Lab: success
columnar-engine C++ No published release Autonomous Engineering Lab: success
lsm-tree C++ No published release Autonomous Engineering Lab: success
counterexample CSS v1.0.0 OpenSSF Scorecard: success
portfolio-ops Python No published release Achievement Scout: success

This block is regenerated only when GitHub's repository, release, or workflow data changes.

↑ ↑ ↓ ↓ ← → ← → B A

Secret level: isolating N1 leaves two leaders in two terms; healing makes N1 step down, rolls back its conflicting entries, and unlocks an achievement.

That was a cartoon. FAULTLINE runs the real C++17 Raft engine in your browser, so you can isolate the leader yourself and watch the safety check decide.

Open channel

aaryansp26@gmail.com · Systems résumé · Systems Observatory · LinkedIn

Engineering OS · Proof tours · PROOFGRAPH · Evidence manifest

UGA computer science, December 2026. Industrial data engineering at MP Equipment. Every system above is MIT licensed: clone one, run the benchmark, and inspect the regime where it breaks.

Thanks for playing: a small robot walks the commit log and zaps split-brain, stale-read, lost-write and off-by-one bugs, turning each one into a test.

All profile visuals are first-party SVGs generated in this repository; the play layer comes from scripts/arcade.py. Every animation is plain CSS and stops under prefers-reduced-motion. Psst: there is a secret level above the contact links.

Pinned Loading

  1. annlite annlite Public

    HNSW vector index from scratch in C++17: SIMD distance kernels, Python bindings, HTTP service, benchmarked against FAISS on recall-vs-latency

    C++

  2. counterexample counterexample Public

    Public engineering failure museum: 12 reproducible counterexamples, exact commits, constrained intake, and signed receipts.

    CSS

  3. edgar-mcp edgar-mcp Public

    MCP server for SEC EDGAR filings and XBRL financial facts

    Python

  4. raft-mvcc raft-mvcc Public

    C++17 Raft + serializable MVCC with linearizability checking; run FAULTLINE live in WebAssembly.

    C++

  5. tensorforge-webgpu tensorforge-webgpu Public

    Typed tensor IR, optimization passes, liveness-aware allocation, generated WGSL, and a live WebGPU correctness lab.

    TypeScript

  6. track-fusion track-fusion Public

    Multi-target tracking from the filter up: IMM estimator, JPDA data association, sequential track scoring, and OSPA evaluation that reports where each approach loses

    Python