{
  "schema_version": 1,
  "kind": "public-portfolio-evidence",
  "generated_from": "src/data/portfolio.ts",
  "scope": "Public project, work-history, and on-page evidence links",
  "boundaries": [
    "Entries link to public artifacts or named sections of this profile; they do not assert access to private systems.",
    "Synthetic fixtures, projections, and controlled local measurements remain labelled in their source evidence.",
    "Production-scale operation, enterprise adoption, and production SLOs are not implied where the linked artifact does not demonstrate them."
  ],
  "proof": [
    {
      "id": "proof-live-demo",
      "kind": "interactive",
      "label": "live demo",
      "statement": "The mission readiness dashboard, running in your browser right now.",
      "href": "https://wafflebits.github.io/readiness-control-tower/"
    },
    {
      "id": "proof-gpu-benchmark-chart",
      "kind": "measured",
      "label": "GPU benchmark chart",
      "statement": "Real RTX 5070 Ti results: my kernels vs PyTorch, chart and raw data in the repo.",
      "href": "https://github.com/WaffleBits/triton-kernel-lab#readme"
    },
    {
      "id": "proof-matching-engine-benchmark",
      "kind": "measured",
      "label": "matching-engine benchmark",
      "statement": "312,000 orders per second, measured and committed with the machine specs.",
      "href": "https://github.com/WaffleBits/market-microstructure-engine#readme"
    },
    {
      "id": "proof-replayable-test-runs",
      "kind": "deterministic",
      "label": "replayable test runs",
      "statement": "Scheduler traces that replay exactly, so any failure can be reproduced.",
      "href": "https://github.com/WaffleBits/deterministic-inference-scheduler"
    },
    {
      "id": "proof-observability-configs",
      "kind": "operational",
      "label": "observability configs",
      "statement": "Grafana dashboards, Prometheus metrics, and incident runbooks in the gateway repo.",
      "href": "https://github.com/WaffleBits/secure-gpu-inference-gateway"
    },
    {
      "id": "proof-restart-reconciliation",
      "kind": "verified",
      "label": "restart reconciliation",
      "statement": "Verified HTTPS CI runs two authenticated agents with an explicit CA bundle, then recovers a completed shard across an agent restart and both shards across a coordinator restart without duplicate target work; partial or multi-host recovery is not claimed.",
      "href": "https://github.com/WaffleBits/triton-inference-benchmark#coordinate-authenticated-agents"
    },
    {
      "id": "proof-controlled-readiness",
      "kind": "verified",
      "label": "controlled readiness",
      "statement": "A real delayed-readiness fixture starts a local service without a shell, measures process launch to HTTP-200 readiness, runs the benchmark, and keeps command and endpoint values out of the report.",
      "href": "https://github.com/WaffleBits/triton-inference-benchmark/blob/main/lifecycle_qualification.py"
    },
    {
      "id": "proof-lifecycle-cancellation",
      "kind": "verified",
      "label": "lifecycle cancellation",
      "statement": "A deterministic CLI fixture interrupts a held-open local benchmark request with SIGINT and SIGTERM, verifies benchmark-child and service-group cleanup, requires exit status 130/143, and checks that no success report is written. This is single-host process-cleanup evidence, not remote cancellation or production recovery.",
      "href": "https://github.com/WaffleBits/triton-inference-benchmark/blob/main/tests/run_lifecycle_signal_fixture.py"
    },
    {
      "id": "proof-qualification-provenance",
      "kind": "verified",
      "label": "qualification provenance",
      "statement": "A content-addressed manifest re-derives a saved trend report and verifies ordered run bytes without storing paths, prompts, endpoints, or credentials.",
      "href": "https://github.com/WaffleBits/triton-inference-benchmark#create-a-content-addressed-qualification-manifest"
    },
    {
      "id": "proof-spending-limits",
      "kind": "operational",
      "label": "spending limits",
      "statement": "Optional Redis-backed limits that stay correct across many servers at once, without storing who made the request.",
      "href": "https://github.com/WaffleBits/secure-gpu-inference-gateway/blob/main/gateway/rate_limit.py"
    },
    {
      "id": "proof-backend-probe",
      "kind": "measured",
      "label": "backend probe",
      "statement": "A short live check against the model endpoint, recording how often it succeeded and how fast.",
      "href": "https://github.com/WaffleBits/secure-gpu-inference-gateway/blob/main/artifacts/backend-probe-evidence.json"
    },
    {
      "id": "proof-telemetry-correlation",
      "kind": "operational",
      "label": "telemetry correlation",
      "statement": "One snapshot tying request outcomes, token usage, and latency together.",
      "href": "https://github.com/WaffleBits/secure-gpu-inference-gateway/blob/main/artifacts/telemetry-correlation-evidence.json"
    },
    {
      "id": "proof-deployment-safety",
      "kind": "verified",
      "label": "deployment safety",
      "statement": "Automated checks that scan the container, list every dependency, and test the locked-down runtime.",
      "href": "https://github.com/WaffleBits/secure-gpu-inference-gateway/blob/main/.github/workflows/supply-chain.yml"
    },
    {
      "id": "proof-failure-mode-drill",
      "kind": "verified",
      "label": "failure-mode drill",
      "statement": "A synthetic gateway drill covers backend latency spikes, backend error bursts, queue saturation, and audit-sink backpressure with explicit detection, mitigation, rollback, and recovery thresholds. It excludes request bodies, identities, secrets, and production logs; it does not claim production resilience.",
      "href": "https://github.com/WaffleBits/secure-gpu-inference-gateway/blob/main/artifacts/resilience-drill-evidence.json"
    },
    {
      "id": "proof-capacity-plan",
      "kind": "verified",
      "label": "capacity plan",
      "statement": "A checked synthetic capacity artifact reports per-model request and input-token cost estimates, observed p95 latency and utilization, and policy envelopes. It excludes request bodies, identities, secrets, and production logs; it does not claim production capacity.",
      "href": "https://github.com/WaffleBits/secure-gpu-inference-gateway/blob/main/artifacts/capacity-plan-evidence.json"
    },
    {
      "id": "proof-stream-usage-validation",
      "kind": "verified",
      "label": "stream usage validation",
      "statement": "Tests reject boolean, negative, fractional, and string completion-token counts, distinguish malformed from missing usage, and omit server-reported token throughput unless every successful stream has a valid count. Text chunks and output bytes are not treated as tokens; the evidence is a deterministic local protocol test, not a production measurement.",
      "href": "https://github.com/WaffleBits/triton-inference-benchmark/blob/main/tests/test_openai_streaming.py"
    },
    {
      "id": "proof-energy-projection-model",
      "kind": "research",
      "label": "energy projection model",
      "statement": "The analog-chip cost model and its projected 68% energy savings, math included.",
      "href": "https://github.com/WaffleBits/heterocore-compiler"
    }
  ],
  "capabilities": [
    {
      "code": "AI",
      "title": "Run and measure AI serving",
      "statement": "Stand up model serving, then prove how fast, reliable, and expensive it is.",
      "topics": [
        "latency percentiles",
        "throughput",
        "cost per request",
        "GPU kernels",
        "benchmarking"
      ],
      "href": "https://github.com/WaffleBits/triton-inference-benchmark",
      "link_label": "Benchmark evidence",
      "link_scope": "public-artifact"
    },
    {
      "code": "SEC",
      "title": "Lock systems down",
      "statement": "Control who gets access, cap what they can spend, and keep records that hold up.",
      "topics": [
        "access control",
        "token budgets",
        "audit trails",
        "threat modeling",
        "incident response"
      ],
      "href": "https://github.com/WaffleBits/secure-gpu-inference-gateway",
      "link_label": "Gateway controls",
      "link_scope": "public-artifact"
    },
    {
      "code": "OBS",
      "title": "Make systems observable",
      "statement": "If something breaks at 3am, the dashboards already know what and why.",
      "topics": [
        "Prometheus",
        "Grafana",
        "OpenTelemetry",
        "release gates",
        "runbooks"
      ],
      "href": "https://github.com/WaffleBits/secure-gpu-inference-gateway#evidence-artifacts",
      "link_label": "Telemetry artifacts",
      "link_scope": "public-artifact"
    },
    {
      "code": "SYS",
      "title": "Build low-level systems",
      "statement": "Write the fast, correct core in Rust or C++ and prove it behaves with tests and replays.",
      "topics": [
        "Rust",
        "C++20",
        "deterministic replay",
        "cross-language testing"
      ],
      "href": "https://github.com/WaffleBits/deterministic-inference-scheduler",
      "link_label": "Runtime evidence",
      "link_scope": "public-artifact"
    },
    {
      "code": "OPS",
      "title": "Operate under pressure",
      "statement": "My day job: enterprise cyber defense for the Air Force, with a TS/SCI clearance.",
      "topics": [
        "vulnerability management",
        "compliance",
        "enterprise tooling"
      ],
      "href": "#experience",
      "link_label": "Experience record",
      "link_scope": "profile-section"
    },
    {
      "code": "LNG",
      "title": "Daily tools",
      "statement": "The stack I actually work in.",
      "topics": [
        "Python",
        "Rust",
        "TypeScript",
        "C++",
        "SQL",
        "Linux",
        "Docker",
        "Kubernetes",
        "Git"
      ],
      "href": "#work",
      "link_label": "Selected work",
      "link_scope": "profile-section"
    }
  ],
  "projects": [
    {
      "id": "CF-01",
      "roles": [
        "security",
        "ai"
      ],
      "title": "Secure GPU Inference Gateway",
      "summary": "A security checkpoint for AI models. It decides who may use which model, enforces spending limits, writes an audit record for every request, and exports live health metrics.",
      "tags": [
        "Security",
        "AI Serving",
        "Python"
      ],
      "problem": "Model-serving paths are high-value and under-governed: who may call which model, at what budget, with what audit trail?",
      "system": "Authenticated gateway enforcing role-based access and reason-for-access policy, with request and token budgets plus an optional Redis-backed atomic limiter path.",
      "evidence": "Audit logs, trace export, Grafana dashboards, and incident runbooks. A checked synthetic resilience drill covers backend latency spikes, backend error bursts, queue saturation, and audit-sink backpressure with explicit detection, mitigation, rollback, and recovery thresholds. A checked synthetic capacity plan reports per-model request and input-token cost estimates, observed p95 latency and utilization, and policy request/token envelopes; it excludes request bodies, identities, secrets, and production logs and does not claim production capacity. Release gates cover shadow, canary, staged rollout, and rollback. Supply-chain checks in CI scan the container image and pin every dependency.",
      "impact": "A defensible, observable model-access boundary with a clear path to policy-as-code and external authorization.",
      "href": "https://github.com/WaffleBits/secure-gpu-inference-gateway"
    },
    {
      "id": "CF-02",
      "roles": [
        "ai"
      ],
      "title": "Deterministic Inference Scheduler",
      "summary": "The traffic controller inside an AI server, rebuilt in Rust so every run is exactly repeatable. Repeatable runs mean failures can be reproduced and new versions can be tested safely.",
      "tags": [
        "Rust",
        "Reliability"
      ],
      "problem": "Serving reliability breaks silently: non-deterministic batching and cache admission make regressions hard to catch before promotion.",
      "system": "Continuous batching with stable priority ordering, paged KV-cache admission, and round-robin decode progress, mirrored against vLLM/SGLang-style observations.",
      "evidence": "Deterministic replay fingerprints, workload-pressure summaries, capacity envelopes, latency checks, and structured hold/rollback triage.",
      "impact": "Promote / hold / rollback decisions backed by reproducible evidence instead of vibes.",
      "href": "https://github.com/WaffleBits/deterministic-inference-scheduler"
    },
    {
      "id": "CF-03",
      "roles": [
        "ai"
      ],
      "title": "Triton Kernel Lab",
      "summary": "Hand-tuned GPU code for two core AI operations, measured honestly against PyTorch's best. Up to 2.2x faster on real hardware, with the charts and raw data published.",
      "tags": [
        "GPU",
        "Performance"
      ],
      "problem": "Kernel-level performance claims are worthless without correctness oracles and controlled measurement.",
      "system": "Fused RMSNorm and autotuned SwiGLU kernels in OpenAI Triton, validated against FP32 reference implementations with explicit cache control.",
      "evidence": "Raw p50/p95/p99 latency distributions from an RTX 5070 Ti, torch.compile baselines, machine-readable reports, and a regression gate in CI.",
      "impact": "Performance claims a reviewer can check: measured, correctness-gated, and regression-protected.",
      "href": "https://github.com/WaffleBits/triton-kernel-lab",
      "chart": "https://github.com/WaffleBits/triton-kernel-lab#readme"
    },
    {
      "id": "CF-04",
      "roles": [
        "ai"
      ],
      "title": "Inference Load Benchmark",
      "summary": "A stress-testing tool for AI serving with authenticated benchmark agents, plus a controlled local lifecycle qualifier that measures readiness before benchmark work. Its deterministic CI fixtures cover verified HTTPS with an explicit CA bundle, bounded restart reconciliation, and no duplicate target work after one coordinator or agent response is lost.",
      "tags": [
        "Benchmarking",
        "Cost"
      ],
      "problem": "Serving decisions need repeatable numbers: latency, throughput, failure accounting, and cost-to-serve.",
      "system": "Repeatable load-generation harness for NVIDIA Triton-style and OpenAI-compatible serving with authenticated coordinator-agent execution, an opt-in TLS 1.2+ agent listener, explicit CA verification, HMAC-bound coordinator state, exact-request status reconciliation, bounded in-memory and SQLite agent result recovery, NTP-style clock sampling, conservative time bounds, open-loop pacing, privacy-safe aggregate gates, and a shell-free local service-lifecycle wrapper that measures process launch to an explicit health response before invoking the benchmark.",
      "evidence": "The TLS fixture generates an ephemeral certificate, runs two authenticated HTTPS agents with an explicit CA bundle, rejects an invalid bearer key and an untrusted CA before target work, and checks redacted JSON/Prometheus artifacts. The published benchmark_report.py compares ordered saved JSON runs with explicit p95-latency, throughput, success-rate, and retry-amplification gates while omitting endpoints and prompts. The lifecycle qualification fixture starts a delayed local service, measures process-launch to HTTP-200 readiness, runs an OpenAI-compatible request path, terminates the service, and omits the command, health URL, source path, prompt, and authorization data from its JSON/Markdown projection. Separate fixtures recover one completed shard across an agent restart and both shards after coordinator restart; the synthetic target remains at exactly eight requests with zero duplicates. The public qualification_manifest.py re-derives the saved trend report and verifies exact ordered input bytes through SHA-256 digests without serializing paths, prompts, endpoints, credentials, outputs, or raw telemetry. OpenAI-compatible streaming completion-token usage accepts only exact non-negative integers; malformed and missing usage are reported separately; server-reported output-token throughput is omitted unless every successful stream has a valid count. Text chunks and output bytes remain separate from token counts. The local lifecycle wrapper also handles SIGINT/SIGTERM during a held-open benchmark request: deterministic fixtures verify benchmark-child and service-group cleanup, exit status 130/143, and no success report. This is single-host process-cleanup evidence, not remote-agent cancellation or recovery. It does not establish partial-workflow continuation, interrupted remote-agent recovery, multi-host behavior, model cold-start time, or production inference measurements.",
      "impact": "Capacity, reliability, and cost tradeoffs expressed as numbers a reviewer can check.",
      "href": "https://github.com/WaffleBits/triton-inference-benchmark"
    },
    {
      "id": "CF-05",
      "roles": [
        "mission"
      ],
      "title": "Readiness Control Tower",
      "summary": "A dashboard that tells a military commander why readiness is dropping and what to fix first, built on synthetic data. Click the live demo and try the what-if controls yourself.",
      "tags": [
        "Mission",
        "Live Demo"
      ],
      "problem": "Operators drown in raw sortie, maintenance, supply, and outage data with no path to root cause under pressure.",
      "system": "FastAPI + React platform with root-cause scoring, what-if analysis, and command-facing workflows over synthetic operational data.",
      "evidence": "Prioritized recommendations, Dockerized deployment, test coverage, and a static live demo on GitHub Pages.",
      "impact": "Ambiguous operational data becomes a decision surface that helps people act under uncertainty.",
      "href": "https://github.com/WaffleBits/readiness-control-tower",
      "demo": "https://wafflebits.github.io/readiness-control-tower/"
    },
    {
      "id": "CF-06",
      "roles": [
        "ai"
      ],
      "title": "HeteroCore Compiler",
      "summary": "Research asking: what if part of an AI chip were analog instead of digital? The cost model projects up to 68% energy savings, clearly labeled as projections, with the full math published.",
      "tags": [
        "Hardware",
        "Research"
      ],
      "problem": "Analog/digital accelerator tradeoffs are opaque: projections, simulations, and measurements get conflated.",
      "system": "ONNX compiler and analytical cost model for mixed analog-digital inference, linked to analog simulation, memory-hierarchy analysis, RTL, and an FPGA prototype.",
      "evidence": "A versioned plan that separates projections, simulations, synthesis outputs, and future board measurements.",
      "impact": "Honest hardware/software co-design analysis with explicit simulation-versus-measurement boundaries.",
      "href": "https://github.com/WaffleBits/heterocore-compiler"
    },
    {
      "id": "CF-07",
      "roles": [
        "mission",
        "ai"
      ],
      "title": "Market Microstructure Engine",
      "summary": "A stock-exchange matching engine built for speed and correctness: 300,000+ orders per second measured, with a Python model double-checking every answer the C++ core gives.",
      "tags": [
        "C++",
        "Low Latency"
      ],
      "problem": "Matching logic is unforgiving: a single ordering bug corrupts every downstream fill.",
      "system": "Python and C++20 matching engines covering price-time priority, partial fills, market orders, and cancellations.",
      "evidence": "Measured benchmark committed to the repo (312K orders/s, p50 1.7 microseconds on a Ryzen 9800X3D), cross-language parity checks, and native tests.",
      "impact": "A benchmarkable, provably-consistent deterministic core with an independent oracle guarding it.",
      "href": "https://github.com/WaffleBits/market-microstructure-engine"
    }
  ]
}