Disability-led research software from Fredericton, NB, Canada.
Reproducible telemetry anomaly detection, C++/HLS research prototypes, RAG guardrails, and offline evidence- and claims-audit tools, with an evidence tag on every number.

sparkainlpx.xyz · LinkedIn · ·
Founder: Jean-François Brisson · Français (langue maternelle) · English (fluent)

Open to work: applied AI research and research-software roles, remote within Canada or on site in Fredericton. Contact details are below.

Spark AI NLP is a disability-led research-software company based in Fredericton, New Brunswick, Canada. The public repositories are research prototypes: a reproducible telemetry anomaly-detection benchmark, the OES-32/OES-512 residual-latch family, C++/HLS kernels, a source-grounded RAG guardrail, two offline audit tools (Evidence Passport for run-level evidence, Quantum Claims Evidence Passport for public claims against their sources), and small offline tools for provenance and consent-first data sharing. They are not hardware products, field products, or medical products, and they do not report results on quantum hardware.

Repository	What it is	Evidence status	Archive
oes-resilience · v0.5.0	Open, reproducible benchmark for 512-channel telemetry anomaly detection with the transparent one-line OES32 reference detector; stress suite, replay evaluation, detector plugin API, and a preregistered NASA SMAP/MSL comparison	Baseline SYNTHETIC (byte-reproduced in CI). On SMAP/MSL under the locked protocol, OES32 did not meet its success criterion
oes32-hls · v0.3.0	C++ HLS prototype of OES-32 triage: streaming AXI4-Stream kernel, pybind11 bindings, and pytest + Hypothesis tests comparing it bit for bit with a Python reference model	g++ testbenches and Python tests run in CI on SYNTHETIC stimuli; FPGA synthesis UNRUN (ZCU111 is a TARGET)
oes512q-latch · v0.2.0	OES-32/OES-512 classical residual latch with a self-calibrated threshold, compared with standard unsupervised detectors on all 47 labelled real-data series in the Numenta Anomaly Benchmark (NAB)	NAB numbers REPORTED (mixed results, reported as measured); unit tests SYNTHETIC; not a quantum code
multi-quantum-oes · v0.1.1	Offline, standard-library-only OES-512 replay-triage workbench with a preregistered stress evaluation, plus an isolated "AI ∩ quantum" toy lab	Stress result is negative and published as such; SYNTHETIC data; exact classical simulation, not a QPU
digital-to-wave-testbench · v0.4.1	Synthetic, dimensionless software testbench for a signed-amplitude sine encoding (noise, phase, CFO, timing; Wilson CIs)	All results SYNTHETIC; no physical validity claimed
evidence-passport · v0.1.0	Offline stdlib audit tool: one experiment-run manifest → static HTML evidence passport + normalized JSON (fail-closed SHA-256); records declared evidence, never re-scores it	Bundled OES-Resilience SMAP/MSL sample keeps the verdict: criterion not met; three further examples labelled SYNTHETIC; sample pages
quantum-claims-passport · v0.1.0	Offline claims audit for scientific communication: classifies public claims (a planned Swiss quantum-computer hub; a modeled 613 THz tubulin frequency band; a rat behavioral study) as announcement, computational model, animal behavioral result, hypothesis, or unsupported inference, with exact-SI conversions and no combined score	Source audit only, snapshot 2026-10-05; makes none of the audited claims; no affiliation with the institutions named; report

Most public research repositories are archived on Zenodo with DOIs (see each README) and grouped in the Spark AI NLP: Research Software Zenodo community.

• Contracts first. Each definition lives in one normative place with executable tests. For OES-32 that is oes32-residual@b77b612 (ADR-001); other implementations are labelled Profile A sidecars and document how they differ.

• Fail-closed. Invalid input is rejected rather than guessed, and so are claims: if something has not been run or measured, the README says so. Negative results are published too.

• Evidence tags on every number.

  Tag	Meaning
  SYNTHETIC	Produced by our own simulation or tests on generated inputs
  REPORTED	Measured by us; not independently reproduced. The README gives the host and conditions, or states plainly that they are unspecified
  TARGET	A design goal; not yet achieved or measured
  UNRUN	Tooling or scripts exist; the run has not been performed

• Reproducible by default. Each README's Quickstart was run from a fresh clone or is marked UNRUN, and repositories with code run their tests in CI on every change.

• Accessibility first. Ask for alternative formats by opening an issue in any repository.

Contribution guidelines, the code of conduct, and the security policy are shared across all repositories: sparkainlp-x/.github.

Each archived repository has a CITATION.cff file (use GitHub's "Cite this repository" button) and a Zenodo concept DOI covering all versions. To refer to exact code, cite the version DOI listed in that repository's README. All records are in the Zenodo community.

Open to applied AI research and research-software roles (remote within Canada, or Fredericton, NB), and to collaboration with research groups, FPGA/decoder engineers, and accessibility-focused partners.

• Website: sparkainlpx.xyz
• LinkedIn: Jean-François Brisson
• ORCID: 0009-0000-9778-5374
• Questions about code: open an issue in the relevant repository