OPEN THE SYSTEMS OBSERVATORY · RÉSUMÉ · EVIDENCE MANIFEST
| enter | what you can verify |
|---|---|
| Verification Theater | inject a published failure into raft-mvcc, edgar-mcp, or track-fusion; replay six deterministic checkpoints; open the source and benchmark drawer |
| 90-second recruiter tour | select Systems, Quant, Defense, or ML Infrastructure and move through a timed, pausable proof narrative |
| Architecture constellation | see how twelve systems connect through one command/source/oracle/limitation contract |
| Evidence desk | ask where a system loses, how it was verified, or how to reproduce it; answers are restricted to the committed manifest |
| Benchmark time machine | inspect the invalid baseline, correction, commit, and current verification command |
Download a role packet: Systems · Quant / Finance · Defense / Autonomy · ML Infrastructure
| recruiter route | enter through | three strongest proofs |
|---|---|---|
| Systems | open route | raft-mvcc · columnar-engine · dst-harness |
| Quant / Finance | open route | edgar-mcp · aad-greeks · lob-market-making |
| Defense / Autonomy | open route | track-fusion · sar-focus · vio-nav |
| ML Infrastructure | open route | annlite · vllm-lite · dist-train |
Currently building: evidence infrastructure that generates this profile, recruiter routes, proof cards, repository artwork, a one-page résumé, and a machine-readable claim-to-benchmark manifest from one source of truth.
| raft-mvcc | edgar-mcp | track-fusion |
|---|---|---|
| Raft + serializable MVCC + linearizability checking | bounded EDGAR tools + cache validation + global pacing | IMM + JPDA + track scoring + OSPA |
| 598 assertions across seeded faults | 138× warm 10-K read; SEC ceiling enforced | 47% lower localization error; failure sweep published |
make test |
make test |
python -m pytest -q |
Start here — reproduce all three flagship results
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 -qThe credential array is generated from the public source itself. The live status badges come directly from each repository's GitHub Actions workflow—no third-party badge service, decorative certifications, or invented awards.
I'm a computer science undergrad at UGA building the layer underneath the model: storage engines, distributed runtimes, retrieval systems, signal processors, and financial-data infrastructure.
Every repository follows the same contract:
Build the mechanism. Attack the assumptions. Measure against a real baseline. Publish where it loses.
That means the output is not a screenshot or a notebook that ran once. It is a reproducible benchmark, a correctness oracle, or a fault campaign that another engineer can clone and challenge.
| system | what is implemented | measured proof |
|---|---|---|
| edgar-mcp (live) | MCP server over SEC EDGAR with identity resolution, filing overflow history, bounded text windows, XBRL discovery, per-host caching, and a global request pacer | 32 tests; warm 10-K read 138× faster; peak request window held at the SEC's 10 req/s ceiling |
| xbrl-normalize (live) | 19 canonical financial line items with period-aware tag selection, restatement handling, derivations, provenance, and accounting-identity checks | 102-company cross-industry benchmark; assets and liabilities at 100% coverage; 96.1% exact identity balance |
| lob-market-making (live) | limit order book with price-time priority and post-only orders, informed-trader flow, three quoting strategies, exact PnL decomposition, and adverse-selection measurement | 36 tests; paired 12-seed comparison; the naive strategy wins on PnL and loses on every risk-adjusted metric at 7× the inventory deviation |
| aad-greeks | tape-based reverse-mode automatic differentiation over Monte Carlo paths, with pricers written once and differentiated rather than hand-derived | 19 tests; Greeks match analytic formulas to 5.6e-16; cost stays 1.7-2.4× the price across a 50× change in input count |
| raft-mvcc | Raft elections, conflict repair and majority commit combined with serializable MVCC, safe-point GC, and P-compositional linearizability checking | 598 assertions across seeded partitions; 11-tick five-node failover in the in-process simulator |
| columnar-engine | C++17 vector-at-a-time engine with null bitmaps, selection vectors, joins, aggregation, TPC-H Q1, and scalar differential oracles | 6,288 assertions; Q1 at 53.1M rows/s; batched filters 1.94× scalar |
| dst-harness | deterministic simulation testing: virtual clock, network and disk, seeded fault injection, Raft's five safety properties, and a Wing-Gong linearizability checker over the client history | 12 of 12 planted Raft defects caught, seeds-to-first-catch 1 to 485; control clean over 2,000 seeds; found 2 real bugs, both written up with reproducing seeds |
| cdcl-sat | CDCL solver in C++17: watched literals with blockers, 1UIP analysis with recursive minimization, VSIDS, Luby restarts, LBD reduction, DRAT proof emission | 1,960 assertions; 180/180 refutations verified by drat-trim, an external checker; 300/300 verdicts agreeing with CaDiCaL 3.0.1 |
| query-planner | histograms, MCV lists and a from-scratch HyperLogLog feeding System R estimation; bushy DP, left-deep DP and greedy search; a vectorised executor so plans can be timed | q-errors to 76× cost 1.00× runtime against a perfect-information oracle, while left-deep search costs 2.15×; cost model rank correlation 0.77 |
| hotstuff-bft | chained HotStuff driven by an adversary that equivocates, withholds, reorders and forges certificates, with checkers scored by a mutation catalogue | safety clean at and below f and broken immediately above it at both n=4 and n=7; 5/9 broken rules caught with 0 false positives; 3 real bugs found, each with a reproducing seed |
| ptq-budget | RTN, GPTQ and AWQ against an exact fp32 reference on a corpus whose optimal perplexity is known in closed form | GPTQ halves perplexity damage while sitting 39% further from the original weights — weight error picks the worse method; granularity alone swings 2-bit damage 239.5% → 30.9% |
| sampling-profiler | SIGPROF sampling with stack aggregation, folded output, and an error analysis scored against workloads of known CPU share | detectability floor derived and verified as n = 4(1−p)/p; publishes the two regimes where it is confidently wrong — 66.3% reported as 10.8%, and a busy thread at 0.0% |
Only results that reduce to a ratio over a named baseline appear above. Throughput, utilisation and drift are not comparable quantities, so putting them on one axis without a baseline would be decoration dressed as evidence — the conformance rates, crash campaigns and coverage numbers stay in the tables where the units survive.
| project | evidence |
|---|---|
| edgar-mcp | EDGAR filing/facts tools; cache validation; iXBRL cleanup; context-safe windows; rate-limit enforcement |
| xbrl-normalize | comparable statements across 102 filers; provenance on every value; missing is never silently zero |
| lob-market-making | post-only matching, informed flow, exact spread-capture/inventory split; two bugs that flattered the results are documented in DESIGN.md |
| aad-greeks (live) | adjoint Greeks at constant cost in the input count; publishes the payoff where pathwise AAD returns a delta of exactly zero and is silently wrong |
| project | evidence |
|---|---|
| lob-market-making | price-time priority, post-only orders, informed flow, exact PnL decomposition, and a paired strategy comparison across 12 seeds |
| aad-greeks | reverse-mode AD over Monte Carlo paths; Greeks match analytic formulas to 5.6e-16 while cost stays 1.7–2.4× the price across a 50× change in input count |
| raft-mvcc | consensus + serializable snapshots + machine-checkable histories |
| columnar-engine | vectorized execution checked against scalar oracles |
| lsm-tree | WAL, memtables, SSTables, Bloom filters, range scans, crash-safe compaction; 4,064 assertions |
| wal-recovery | CRC32 records, damaged-tail repair, atomic replay; 100 crash trials with zero acknowledged writes lost |
| query-planner | join ordering, cost model and cardinality estimation scored against a perfect-information oracle; reports that the search space matters 20× more than the estimates |
| project | evidence |
|---|---|
| dst-harness | deterministic simulation with seeded partitions, crashes and torn writes; scored by seeds-to-first-catch over 12 planted Raft defects, and it found two real ones in its own target |
| cdcl-sat | every UNSAT answer emitted as a DRAT refutation and verified by drat-trim, not by me; publishes the ablation where its own restart strategy is worth nothing |
| sampling-profiler | scores a profiler against known ground truth and reports where it cannot see: a C call at 6× underattribution and a worker thread at zero |
| hotstuff-bft | proves the adversary is strong enough before making any safety claim — the same harness must break safety above f, and does; four surviving mutants are published rather than dropped |
| project | evidence |
|---|---|
| ptq-budget | weight-only INT8/4/3/2 with an exact oracle; publishes that its own headline metric, weight error, ranks the methods backwards |
| vllm-lite | paged KV cache, continuous batching, prefix caching, speculative decoding; 94% vs 21% KV utilization |
| annlite | HNSW + SIMD + Python bindings; FAISS parity at 0.99 recall and 1.83× faster at 0.999 |
| dist-train | ring all-reduce from send/recv; 58.7 vs 234.9 MB/worker at eight workers |
| rag-eval | hybrid retrieval, reranking, hallucination metrics, and CI gates; SciFact 0.6643 vs 0.665 published BM25 |
| feature-store | point-in-time joins and streaming materialization; leakage measured at +0.059 AUC |
| grammar-decode | JSON Schema → character automaton → cached token mask; 100% vs 0% conformance baseline |
| agent-harness | five-layer sandbox and graded benchmark; 58 tests, including 26 escape attempts |
| codebase-qa | auth, rate limits, durable jobs, cost tracking, and tracing—the prototype-to-service gap |
| project | evidence |
|---|---|
| sdr-receiver | QPSK, RRC, acquisition, soft decisions, LDPC; 0 errors in 21,600 bits at 4–8 dB after the waterfall |
| sar-focus | pulse compression, backprojection, PGA autofocus; resolution within 0.5% of theory |
| track-fusion | IMM + JPDA + Wald scoring + OSPA; reports the manoeuvre regime where IMM wins and straight-line regime where it costs |
| vio-nav | MSCKF, SO(3) preintegration, null-space feature marginalisation; 51.9× lower drift than inertial dead reckoning |
Four of these publish a live page built in the same repository — no CDN, no third-party script, and every figure traceable to the benchmark that produced it.
| page | what you can drive |
|---|---|
| aad-greeks | drag the bump size and watch a finite difference find its error floor and climb back out of it; drag payoff smoothing and watch an adjoint delta go from exactly zero to correct |
| lob-market-making | a live order book getting picked off, with sliders for spread, inventory skew and toxic flow; re-sort the strategy table and watch the ranking invert |
| xbrl-normalize | step the tag resolver through five filers and watch it reject stale tags; toggle mezzanine equity and watch the balance sheet stop balancing |
| edgar-mcp | fire a burst of tool calls at a token bucket and a strict pacer, and see which one breaks the limit it enforces |
Regenerated daily by GitHub
Actions. The panel reads the repositories and counts test functions from
source; no third-party stats card, hand-edited total, or page-load API call is
involved. Parametrized suites and looped invariants expand into many more checks
than the function count—for example, raft-mvcc alone executes 598
assertions.
Every item appears in public source, a build manifest, a CI workflow, or a reproducible verification path in this portfolio. The matrix is generated in this repository and does not depend on a third-party badge service.
SYSTEMS C++17 · Python · Raft · MVCC · WAL · LSM/SSTables · SIMD · POSIX
ML INFRA PyTorch · distributed collectives · HNSW · BM25 · reranking · evals
FINANCE SEC EDGAR · XBRL · accounting identities · provenance · MCP
MARKETS limit order books · price-time priority · adverse selection · Avellaneda-Stoikov
QUANT adjoint AD · pathwise Greeks · Black-Scholes · Monte Carlo · payoff smoothing
SIGNAL QPSK · RRC · LDPC · SAR · Kalman/IMM · JPDA · OSPA
ESTIMATE IMU preintegration · SO(3) · MSCKF · ATE/RPE · triangulation
SERVICES FastAPI · SQLite/Postgres · durable queues · scrypt · tracing
INDUSTRY Power BI · Node-RED · CtrlX Core · ERP automation
Day job: automation and analytics at MP Equipment—dashboards, ERP automation, and AR/AI for industrial food processing.
aaryansp26@gmail.com · one-page systems résumé · Systems Observatory
Every system above is MIT licensed. Clone one, run the benchmark, and inspect the regime where it breaks.
Hero, topology, and telemetry are generated in this repository. Both
themes are first-party SVGs and respect prefers-reduced-motion.

