{ "version": 2, "generated": "2026-08-09", "owner": { "name": "Aaryan Patel", "email": "aaryansp26@gmail.com", "github": "https://github.com/asp53826", "linkedin": "https://www.linkedin.com/in/aaryanpatelsystems/", "site": "https://asp53826.github.io", "role": "Systems and ML Infrastructure Engineer", "education": "Computer Science undergraduate at the University of Georgia", "work": "Automation and analytics at MP Equipment" }, "thesis": "I build systems that survive faults, bad data, and dishonest benchmarks.", "contract": "Build the mechanism. Attack the assumptions. Measure against a real baseline. Publish where it loses.", "health": { "measuredSystems": 24, "publicRepositories": 34, "testFunctions": 815, "sourceFiles": 501, "languages": 8, "liveLabs": 5, "trackedWorkflows": 12, "lastVerified": "2026-08-09" }, "current": { "title": "TensorForge visual WebGPU compiler", "repo": "tensorforge-webgpu", "status": "live", "summary": "An inspectable tensor compiler that rewrites a typed graph, generates WGSL, executes it on the visitor's GPU, and verifies the result against a CPU oracle.", "proof": "Nine tests plus browser QA cover the lowering pipeline, memory slots, generated kernels, and numerical agreement." }, "experience": { "verifiedOn": "2026-08-09", "disclosure": "Interactive replays are deterministic narrative checkpoints derived from published tests and benchmark tables. They are not live production telemetry.", "flagships": [ { "projectId": "raft-mvcc", "commit": "dbd5398", "benchmarkCommand": "make clean test benchmark", "sourceFiles": [ "tests/test_raft_mvcc.cpp", "bench/benchmark.cpp", "src/raft.cpp", "src/linearizability.cpp" ], "scenarioLabel": "Leader isolation and log repair", "scenarioSummary": "Isolate the leader, allow an unsafe proposal on the minority side, elect a majority leader, then heal the partition.", "invariant": "A minority leader may append locally but cannot advance the committed index.", "result": "The majority fails over in 11 logical ticks and the old leader's conflicting suffix is repaired after healing.", "events": [ { "time": "T+00", "label": "Five nodes agree", "detail": "Node 1 leads a healthy five-node cluster; the committed prefix matches on every node.", "state": "healthy" }, { "time": "T+01", "label": "Partition injected", "detail": "Node 1 is isolated from the four-node majority. Messages crossing the cut are deterministically dropped.", "state": "fault" }, { "time": "T+03", "label": "Minority proposal", "detail": "The isolated leader accepts an uncommitted local proposal but cannot obtain a quorum.", "state": "contained" }, { "time": "T+07", "label": "Fresh candidate wins", "detail": "The majority elects a leader using the term and last-log freshness checks.", "state": "recovering" }, { "time": "T+11", "label": "Majority commits", "detail": "The new leader commits on the majority while the isolated leader remains unable to commit.", "state": "verified" }, { "time": "HEAL", "label": "Conflict repaired", "detail": "Healing exposes the stale suffix; nextIndex backtracking truncates and replaces it.", "state": "verified" } ], "history": [ { "label": "Mechanism release", "commit": "dbd5398", "metric": "598 assertions", "note": "Targeted invariants plus a seeded failover campaign." }, { "label": "Published failover", "commit": "dbd5398", "metric": "11 logical ticks", "note": "Five-node leader isolation and majority re-election." }, { "label": "Verification surface", "commit": "b4eef31", "metric": "CI badge added", "note": "Repository metadata now exposes the same correctness path." } ], "postmortem": { "failure": "An isolated leader can still accept a local append, which looks successful if commit state is ignored.", "signal": "The test tracks commit advancement separately from proposal acceptance.", "correction": "Only a current-term entry replicated to a majority can advance the leader's commit index.", "lesson": "Availability-looking behavior is not safety evidence; the invariant must be observed at the commit boundary." } }, { "projectId": "edgar-mcp", "commit": "825052d", "benchmarkCommand": "make test && make bench", "sourceFiles": [ "src/edgar_mcp/client.py", "tests/test_client.py", "tests/test_tools.py", "README.md" ], "scenarioLabel": "Burst control and truthful cache hits", "scenarioSummary": "Drive concurrent EDGAR reads through the original burst-prone design, then replay the strict pacer and host-aware cache correction.", "invariant": "Request grants stay paced below the SEC ceiling, and a cache hit transfers no response body.", "result": "The corrected client sustains 9.2 requests per second and serves a warm 10-K in 1.1 ms with zero downloaded megabytes.", "events": [ { "time": "T+00", "label": "Cold filing read", "detail": "The client fetches a 1.52 MB 10-K and stores the immutable archive document.", "state": "healthy" }, { "time": "T+01", "label": "Burst model challenged", "detail": "The original full token bucket places 19 grants inside a one-second window.", "state": "fault" }, { "time": "T+02", "label": "Strict pacing enabled", "detail": "A global monotonic pacer spaces grants and removes startup burst capacity.", "state": "contained" }, { "time": "T+03", "label": "False cache hit exposed", "detail": "A reported hit still transfers 3.75 MB because the host cannot validate it as assumed.", "state": "fault" }, { "time": "T+04", "label": "Freshness split by host", "detail": "Immutable archives cache forever; data APIs use TTL or validators that the host actually provides.", "state": "recovering" }, { "time": "T+05", "label": "Warm read verified", "detail": "The 10-K returns in 1.1 ms with 0.00 MB downloaded, a measured 138x improvement.", "state": "verified" } ], "history": [ { "label": "First rate benchmark", "commit": "pre-fix", "metric": "19 req / 1 s", "note": "A full token bucket nearly doubled the intended ceiling." }, { "label": "Strict pacer", "commit": "current", "metric": "9.2 req / s", "note": "Grants are globally spaced with no initial burst." }, { "label": "First cache benchmark", "commit": "pre-fix", "metric": "3.75 MB warm", "note": "The hit counter hid a complete re-download." }, { "label": "Host-aware cache", "commit": "current", "metric": "0.00 MB warm", "note": "Warm 10-K read measures 1.1 ms and 138x speedup." } ], "postmortem": { "failure": "The first implementation looked compliant and cached because its abstractions reported tokens and hits.", "signal": "Wire-level one-second request windows and transferred-byte counters contradicted the abstraction-level metrics.", "correction": "Replace burst capacity with strict pacing and assign cache freshness rules per EDGAR host.", "lesson": "Measure the wire, not the class name: a RateLimiter and Cache do not prove rate limiting or caching." } }, { "projectId": "track-fusion", "commit": "099601b", "benchmarkCommand": "python -m pytest -q && python bench/benchmark.py --quick", "sourceFiles": [ "tf/imm.py", "tf/assoc/jpda.py", "tf/metrics.py", "bench/benchmark.py" ], "scenarioLabel": "Maneuver, crossing, and cardinality failure", "scenarioSummary": "Track a maneuvering target, introduce a crossing under clutter, then compare localization accuracy with full OSPA cost.", "invariant": "Association comparisons hold the tracker constant, and evaluation charges for both localization and target-count errors.", "result": "IMM lowers localization error 47% inside its useful turn-rate band; JPDA improves crossing localization but sustains 2.7x the false tracks.", "events": [ { "time": "T+00", "label": "Straight leg", "detail": "The CV and IMM estimators agree while motion remains close to constant velocity.", "state": "healthy" }, { "time": "T+01", "label": "Turn begins", "detail": "Measurements shift probability toward the coordinated-turn modes in the IMM bank.", "state": "recovering" }, { "time": "T+02", "label": "Localization measured", "detail": "OSPA localization falls from 26.92 for the single CV filter to 14.40 for IMM.", "state": "verified" }, { "time": "T+03", "label": "Tracks cross", "detail": "Two measurements become plausible for both targets under light clutter.", "state": "fault" }, { "time": "T+04", "label": "Soft association wins locally", "detail": "JPDA localization reaches 13.76 versus 14.70 for GNN across the crossing.", "state": "contained" }, { "time": "T+05", "label": "Cardinality exposes the loss", "detail": "JPDA keeps 2.7x the false tracks; total OSPA is 39.55 versus 28.73 for GNN.", "state": "verified" } ], "history": [ { "label": "Single CV baseline", "commit": "099601b", "metric": "26.92 OSPA loc", "note": "Coverage is 0.893 through the maneuver." }, { "label": "IMM bank", "commit": "099601b", "metric": "14.40 OSPA loc", "note": "Coverage rises to 0.960 inside the 3-6 degree/s useful band." }, { "label": "Crossing GNN", "commit": "099601b", "metric": "28.73 total OSPA", "note": "Hard association has slightly worse localization and much lower cardinality error." }, { "label": "Crossing JPDA", "commit": "099601b", "metric": "39.55 total OSPA", "note": "Localization improves while false tracks rise from 0.21 to 0.56 per scan." } ], "postmortem": { "failure": "RMSE and localization-only views make JPDA look like the better crossing strategy.", "signal": "OSPA cardinality cost reveals marginal clutter tracks that soft association keeps alive.", "correction": "Publish localization and cardinality separately, then compare the total under paired seeds.", "lesson": "A better estimator can be a worse tracking system when the metric ignores existence errors." } } ] }, "routes": [ { "id": "systems", "label": "Systems", "short": "SYS", "summary": "Storage, execution, consensus, and fault discovery.", "projects": [ "raft-mvcc", "columnar-engine", "dst-harness" ] }, { "id": "quant", "label": "Quant / Finance", "short": "QNT", "summary": "Financial data, market structure, and differentiable pricing.", "projects": [ "edgar-mcp", "aad-greeks", "lob-market-making" ] }, { "id": "defense", "label": "Defense / Autonomy", "short": "DEF", "summary": "Tracking, radar image formation, and GPS-denied navigation.", "projects": [ "track-fusion", "sar-focus", "vio-nav" ] }, { "id": "ml-infrastructure", "label": "ML Infrastructure", "short": "MLI", "summary": "Retrieval, inference serving, and distributed training.", "projects": [ "annlite", "vllm-lite", "dist-train" ] } ], "projects": [ { "id": "raft-mvcc", "name": "raft-mvcc", "track": "systems", "language": "C++17", "tagline": "Consensus, serializable snapshots, and machine-checkable histories.", "problem": "A replicated database can acknowledge writes while violating safety under partitions, leader changes, or transaction conflicts.", "mechanism": "Raft elections and conflict repair combined with MVCC snapshots, serializable OCC, safe-point GC, and P-compositional linearizability checking.", "attack": "Seeded partitions, stale candidates, minority leaders, conflicting transactions, tombstones, and invalid histories.", "metric": "598", "metricLabel": "assertions across seeded faults", "proof": "Five-node failover completes in 11 simulator ticks while minority leaders cannot commit.", "limitation": "An in-process simulator: no sockets, fsync, snapshots, membership changes, or multi-range transactions.", "command": "git clone https://github.com/asp53826/raft-mvcc && cd raft-mvcc && make test", "repo": "https://github.com/asp53826/raft-mvcc", "methods": [ "linearizability", "fault injection", "differential invariants" ] }, { "id": "columnar-engine", "name": "columnar-engine", "track": "systems", "language": "C++17", "tagline": "Vector-at-a-time analytical execution checked against scalar oracles.", "problem": "Analytical operators can be fast while silently mishandling nulls, selections, joins, or aggregation.", "mechanism": "Null bitmaps, selection vectors, vectorized filters, joins, aggregation, and TPC-H Q1.", "attack": "Every vectorized operator is differentially checked against a scalar reference.", "metric": "53.1M", "metricLabel": "TPC-H Q1 rows per second", "proof": "6,288 assertions; batched filters are 1.94x the scalar implementation.", "limitation": "In-memory engine without persistence, distributed execution, or a SQL frontend.", "command": "git clone https://github.com/asp53826/columnar-engine && cd columnar-engine && make test", "repo": "https://github.com/asp53826/columnar-engine", "methods": [ "scalar oracle", "differential testing", "benchmark" ] }, { "id": "dst-harness", "name": "dst-harness", "track": "systems", "language": "Python", "tagline": "Deterministic simulation testing scored by seeds-to-first-catch.", "problem": "Distributed failures are hard to reproduce when time, network, disk, and scheduling remain nondeterministic.", "mechanism": "Virtual clock, network and disk, seeded fault injection, Raft safety properties, and a Wing-Gong checker.", "attack": "Twelve planted Raft defects plus 2,000 clean control seeds.", "metric": "12 / 12", "metricLabel": "planted defects caught", "proof": "Seeds-to-first-catch range from 1 to 485; the harness also found two real bugs with reproducing seeds.", "limitation": "The target protocol model is deliberately smaller than a production Raft implementation.", "command": "git clone https://github.com/asp53826/dst-harness && cd dst-harness && python3 -m venv .venv && .venv/bin/pip install -e '.[dev]' && .venv/bin/python -m pytest -q", "repo": "https://github.com/asp53826/dst-harness", "methods": [ "deterministic simulation", "mutation testing", "linearizability" ] }, { "id": "edgar-mcp", "name": "edgar-mcp", "track": "quant", "language": "Python", "tagline": "SEC EDGAR tools with bounded context, caching, and global pacing.", "problem": "Filings are large, issuer identities are ambiguous, and naive concurrent clients can violate the SEC request ceiling.", "mechanism": "Identity resolution, filing history, bounded text windows, XBRL discovery, per-host cache validation, and a global request pacer.", "attack": "Cold and warm reads, burst traffic, identity collisions, overflow history, and malformed iXBRL.", "metric": "138x", "metricLabel": "warm 10-K read vs cold fetch", "proof": "32 offline tests; peak request windows stay at the SEC 10 requests-per-second ceiling.", "limitation": "Public EDGAR data only; it is not an investment recommendation or a complete accounting model.", "command": "git clone https://github.com/asp53826/edgar-mcp && cd edgar-mcp && make test", "repo": "https://github.com/asp53826/edgar-mcp", "live": "https://asp53826.github.io/edgar-mcp/", "methods": [ "rate-limit enforcement", "cache benchmark", "fixture tests" ] }, { "id": "aad-greeks", "name": "aad-greeks", "track": "quant", "language": "Python", "tagline": "Reverse-mode automatic differentiation over Monte Carlo paths.", "problem": "Bumping every pricing input scales linearly and finite differences collapse near discontinuous payoffs.", "mechanism": "A tape-based reverse-mode AD engine with pricers written once and differentiated instead of hand-derived.", "attack": "Analytic Black-Scholes formulas, central differences across bump sizes, and discontinuous-payoff failure cases.", "metric": "5.6e-16", "metricLabel": "maximum analytic error", "proof": "19 tests; all Greeks cost 1.7-2.4x the price across a 50x change in input count.", "limitation": "Pathwise derivatives return a wrong zero delta for an unsmoothed digital payoff; the live lab exposes it.", "command": "git clone https://github.com/asp53826/aad-greeks && cd aad-greeks && make test", "repo": "https://github.com/asp53826/aad-greeks", "live": "https://asp53826.github.io/aad-greeks/", "methods": [ "analytic oracle", "finite differences", "complexity benchmark" ] }, { "id": "lob-market-making", "name": "lob-market-making", "track": "quant", "language": "Python", "tagline": "Price-time matching and market making under toxic flow.", "problem": "Raw PnL can make a strategy look strong while inventory exposure and adverse selection carry the risk.", "mechanism": "Limit order book, post-only orders, informed flow, three quoting strategies, and exact PnL decomposition.", "attack": "Paired 12-seed strategy comparison with spread, inventory, and toxicity held explicit.", "metric": "7x", "metricLabel": "naive inventory deviation", "proof": "36 tests; naive quoting wins raw PnL and loses every risk-adjusted metric.", "limitation": "Synthetic order flow without exchange latency, queue-position uncertainty, or real fees.", "command": "git clone https://github.com/asp53826/lob-market-making && cd lob-market-making && make test", "repo": "https://github.com/asp53826/lob-market-making", "live": "https://asp53826.github.io/lob-market-making/", "methods": [ "paired seeds", "PnL attribution", "regime sweep" ] }, { "id": "track-fusion", "name": "track-fusion", "track": "defense", "language": "Python", "tagline": "IMM, JPDA, track scoring, and OSPA from the filter up.", "problem": "Tracking fails across estimation, association, and track existence; RMSE alone hides cardinality failure.", "mechanism": "IMM motion bank, exact JPDA enumeration with bounded fallback, Wald track scoring, and OSPA decomposition.", "attack": "Turn-rate sweeps, crossing targets, clutter, and paired association comparisons.", "metric": "47%", "metricLabel": "localization error reduction", "proof": "67 tests; IMM coverage rises from 0.893 to 0.960, while the published sweep shows where the bank loses.", "limitation": "The useful IMM band is 3-6 degrees per second; at 0 and 12 degrees per second the configured bank loses.", "command": "git clone https://github.com/asp53826/track-fusion && cd track-fusion && python -m pip install -e '.[dev]' && python -m pytest -q", "repo": "https://github.com/asp53826/track-fusion", "methods": [ "OSPA", "paired seeds", "parameter sweep" ] }, { "id": "sar-focus", "name": "sar-focus", "track": "defense", "language": "Python", "tagline": "Image formation checked against analytic resolution and sidelobe theory.", "problem": "A plausible radar image can still have the wrong resolution or impossible sidelobes.", "mechanism": "LFM pulse compression, time-domain backprojection, phase-gradient autofocus, and impulse-response analysis.", "attack": "Independent bandwidth/aperture sweeps and theory checks for resolution and PSLR.", "metric": "<0.5%", "metricLabel": "resolution error vs theory", "proof": "27 tests; the theory oracle exposed interpolation that produced physically impossible sidelobes.", "limitation": "Synthetic phase history with idealized propagation and no measured sensor calibration errors.", "command": "git clone https://github.com/asp53826/sar-focus && cd sar-focus && python -m pip install -e '.[dev]' && python -m pytest -q", "repo": "https://github.com/asp53826/sar-focus", "methods": [ "analytic theory", "double dissociation", "impulse response" ] }, { "id": "vio-nav", "name": "vio-nav", "track": "defense", "language": "Python", "tagline": "MSCKF visual-inertial odometry built from the Lie algebra up.", "problem": "IMU bias integrates into unbounded position drift while a monocular camera cannot observe scale alone.", "mechanism": "SO(3) preintegration, sliding camera-state window, null-space feature marginalization, and ATE/RPE evaluation.", "attack": "Duration sweeps, hover/circle/figure-eight motion, parallax rejection, and dead-reckoning baseline.", "metric": "51.9x", "metricLabel": "lower drift at 40 seconds", "proof": "37 tests; MSCKF stays near four centimeters while dead reckoning reaches 2.106 meters.", "limitation": "At five seconds VIO is worse than inertial-only, and hover makes it 13x worse because depth is unobservable.", "command": "git clone https://github.com/asp53826/vio-nav && cd vio-nav && python -m pip install -e '.[dev]' && python -m pytest -q", "repo": "https://github.com/asp53826/vio-nav", "methods": [ "ground-truth trajectory", "motion sweep", "ATE and RPE" ] }, { "id": "annlite", "name": "annlite", "track": "ml-infrastructure", "language": "C++17 + Python", "tagline": "HNSW with SIMD kernels, Python bindings, and FAISS comparison.", "problem": "Approximate retrieval claims are meaningless without the full recall-versus-latency curve.", "mechanism": "Layered HNSW graph, diversity-aware neighbor selection, SIMD distance kernels, and GIL-free search.", "attack": "Exact ground truth, recall-matched FAISS comparison, and thread-count invariance.", "metric": "1.83x", "metricLabel": "FAISS QPS at 0.999 recall", "proof": "51 tests; at 98.5% recall it performs 26x fewer distance computations than exact scan.", "limitation": "FAISS remains 10-20% faster in the low-recall regime; the crossover appears only at high recall.", "command": "git clone https://github.com/asp53826/annlite && cd annlite && pip install -r requirements-dev.txt && make test", "repo": "https://github.com/asp53826/annlite", "methods": [ "exact recall", "FAISS baseline", "concurrency invariants" ] }, { "id": "vllm-lite", "name": "vllm-lite", "track": "ml-infrastructure", "language": "Python", "tagline": "Paged KV cache and continuous batching around a small model.", "problem": "Static batching strands KV memory and creates head-of-line blocking under variable request lengths.", "mechanism": "Paged cache, continuous batching, chunked prefill, prefix caching, and speculative decoding.", "attack": "Same weights, sampling, traffic distribution, and memory budget across sequential, static, and paged schedulers.", "metric": "94%", "metricLabel": "KV slot utilization vs 21%", "proof": "90 tests; paged scheduling reaches 2.9x sequential throughput and 5.5x better median TTFT.", "limitation": "CPU-scale simulator with a small model; it does not claim production GPU kernel performance.", "command": "git clone https://github.com/asp53826/vllm-lite && cd vllm-lite && pip install -r requirements.txt && pytest", "repo": "https://github.com/asp53826/vllm-lite", "methods": [ "scheduler benchmark", "memory accounting", "token invariance" ] }, { "id": "dist-train", "name": "dist-train", "track": "ml-infrastructure", "language": "Python", "tagline": "Ring all-reduce built from send/recv, then measured where it loses.", "problem": "A one-line collective hides the communication cost and the scale at which latency dominates bandwidth efficiency.", "mechanism": "Ring all-reduce, data parallelism, tensor parallelism, pipeline parallelism, and explicit cost models.", "attack": "Every collective is checked against torch.distributed across worker counts and tensor sizes.", "metric": "58.7 MB", "metricLabel": "per worker at eight workers", "proof": "57 tests; the naive baseline sends 234.9 MB per worker under the same eight-worker case.", "limitation": "Loopback gloo on a 10-core machine; efficiency falls past four workers and small tensors favor naive collectives.", "command": "git clone https://github.com/asp53826/dist-train && cd dist-train && pip install -r requirements.txt && pytest", "repo": "https://github.com/asp53826/dist-train", "methods": [ "torch.distributed oracle", "cost model", "scaling sweep" ] }, { "id": "tensorforge-webgpu", "name": "tensorforge-webgpu", "track": "ml-infrastructure", "language": "TypeScript + WGSL", "tagline": "A visual tensor compiler that lowers live graphs into browser-native GPU kernels.", "problem": "Compiler and GPU portfolio work is difficult to evaluate when the optimization pipeline is hidden behind build logs or requires a local toolchain.", "mechanism": "Typed tensor IR, shape inference, identity canonicalization, operator fusion, output-rooted dead-code elimination, liveness-aware buffer reuse, and three generated WebGPU compute kernels.", "attack": "Deterministic CPU/GPU comparison, incompatible-shape rejection, persistent-slot isolation, fusion toggles, memory-plan tests, and responsive browser QA.", "metric": "8 → 3", "metricLabel": "primitive kernels to GPU dispatches", "proof": "9 tests; the default graph lowers from 13 to 8 nodes and the live GPU result is checked against a seeded CPU oracle.", "limitation": "A narrow dense-f32 MLP experiment: kernels favor readability over autotuning, softmax is not a parallel reduction, and timing is hardware-local queue wall time.", "command": "git clone https://github.com/asp53826/tensorforge-webgpu && cd tensorforge-webgpu && npm ci && npm run check", "repo": "https://github.com/asp53826/tensorforge-webgpu", "live": "https://asp53826.github.io/tensorforge-webgpu/", "methods": [ "CPU oracle", "compiler pass tests", "browser GPU validation" ] } ] }