Reproducible AI & optical-simulation software
Evidence-first engineering · transparent benchmarks · interactive scientific demos
I build research software that makes claims inspectable, rerunnable, and falsifiable: telemetry anomaly detection, evidence auditing, RAG guardrails, digital holography, optical simulation, and interactive scientific tools.
Core rule: a result is useful only if another person can inspect its assumptions, rerun it, and see where it fails.
I am the founder of Spark AI NLP, a disability-led research-software initiative in Fredericton, New Brunswick, Canada. The portfolio spans classical simulation, browser demos, offline audit tools, numerical-contract testing, and C++/HLS prototypes. Each repository states its own evidence status and limitations.
| If you want to… | Start with |
|---|---|
| Assess research integrity | Spark Evidence Lab · Evidence Passport |
| Reproduce a telemetry result | OES Resilience · OES-32 Residual |
| Explore holography and optics | WaveHolo · WaveHolo Telescope |
| Review quantum-inspired engineering | qldpc_decoder_cpp · qec-scaling-bench |
| Test responsible RAG | spark-rag-guardrail |
| See the complete project index | PORTFOLIO.md |
The portfolio is being consolidated around five outcome-oriented research tracks. Each track prioritizes a locked protocol, a strong baseline, uncertainty reporting, and an explicit replication package over additional standalone prototypes.
| Track | Scientific outcome sought | Primary repositories | Next decisive milestone |
|---|---|---|---|
| Telemetry anomaly detection | Determine whether OES-style residual methods add value over strong, transparent baselines on public benchmarks. | oes-resilience, oes32-residual | Publish a locked, preregistered evaluation with confidence intervals, ablations, and a reproducible result archive. |
| Digital holography | Establish which focus and wavefront metrics are reliable under controlled synthetic and real-data conditions. | waveholo, waveholo-telescope | Expand the real-data slope check into a protocol-locked validation study with error bars and failure cases. |
| Quantum-inspired decoding | Turn the qLDPC/HLS scaffold into a measurable software-to-hardware pathway. | qldpc_decoder_cpp, qec-scaling-bench | Complete the BP kernel and HIL verification plan, then report synthesis, resource, latency, and decoder-performance results or explicit blockers. |
| Evidence infrastructure | Make experimental claims mechanically inspectable, replayable, and auditable. | spark-evidence-lab, evidence-passport, claims-layer | Standardize one evidence-passport schema and demonstrate end-to-end replay for at least three independent experiments. |
| Responsible RAG | Quantify when retrieval-grounded refusal improves answer reliability. | spark-rag-guardrail | Run a preregistered benchmark comparing answer quality, hallucination rate, and refusal behaviour against a no-guardrail baseline. |
Decision rule: a track advances only when it produces a rerunnable result, a clear baseline comparison, quantified uncertainty, and an explicit verdict: reproduced, baseline stronger, open question, or exploratory.
| Demo | What it shows | Link |
|---|---|---|
| Wavefront Clinic | Correct Zernike aberrations, inspect the Fraunhofer PSF, and meet the Marechal Strehl criterion in an offline synthetic optics game. | Open demo |
| WaveHolo game | Interactive off-axis digital holography and focus estimation. | Open demo |
| Holographic Thoughts | Accessibility-oriented interactive holographic-interface prototype. | Open demo |
| Spark Evidence Lab | Rerunnable experiment cards, explicit baselines, uncertainty, and negative results. | Open board |
Reproducible experiment cards with explicit baselines and negative results.
The evidence layer for this portfolio. Each card records the claim tested, conditions, baseline comparison, uncertainty, status, source commit, DOI where available, and rerun commands.
- Live board: https://sparkainlp-x.github.io/spark-evidence-lab/
- DOI: 10.5281/zenodo.23403169
- Verdicts: Reproduced, Baseline stronger, Open question, Exploratory
2. WaveHolo
End-to-end off-axis digital holography: record, reconstruct, and estimate focus.
Includes angular-spectrum reconstruction, sub-bin carrier estimation, focus estimation, a browser research sandbox, CI tests, and a real-data benchmark harness.
- Live demo: https://sparkainlp-x.github.io/waveholo/game/
- v0.1.5 DOI: 10.5281/zenodo.23653721
- Reported real-data check: 530 microscope holograms at 6 stage positions; a slope check, not an absolute-focus validation.
Synthetic telescope-pupil holography, Zernike wavefronts, and transparent metric failures.
Benchmarks pupil-plane focus, recovered wavefront agreement, pipeline floor, and noise sensitivity. It reports that phase_variance does not locate focus on telescope-pupil data.
- Interactive game: Wavefront Clinic
- Status: Synthetic only; no real telescope or instrument claim
Telemetry anomaly detection evaluated under locked conditions.
Tests anomaly-detection methods on public telemetry benchmarks, including NASA SMAP/MSL, with preregistered comparisons and transparent reporting.
- DOI: 10.5281/zenodo.23071166
- Published outcome: OES-32 and its locked v2 re-attempt did not meet their success criteria.
Cross-language numerical-contract testing.
A JavaScript port validated against Python-generated contract vectors for bit-exact replay, with pinned source commits and CI regeneration.
Provenance and disclosure infrastructure for AI experiments.
Turns an experiment-run manifest into a static HTML evidence passport and normalized JSON. It records declared evidence and limitations; it does not re-score or validate evidence.
- Sample passports: https://sparkainlp-x.github.io/evidence-passport/
- DOI: 10.5281/zenodo.23165143
| Experiment | Status | Source |
|---|---|---|
| Blind NASA SMAP/MSL test, OES32 v1 vs baselines | Baseline stronger / criterion not met | oes-resilience |
| OES-32-v2 locked re-attempt | Criterion not met | oes-resilience |
| OES-32 residual JS port and live replay | Reproduced | oes32-residual |
| WaveHolo microscope slope check | REPORTED; slope check, not absolute-focus validation | waveholo |
| WaveHolo Telescope pupil reconstruction | Synthetic; negative result for phase_variance |
waveholo-telescope |
| Surface-code memory toy, Stim + PyMatching | Open question | qec-scaling-bench |
- Baselines first. A method is meaningful only relative to a clearly specified alternative.
- Negative results are results. Where a baseline wins, report it.
- Pin the evidence. Claims link to commits, datasets, commands, or DOIs.
- Label uncertainty. Reproduced, exploratory, and open questions are not equivalent.
- Separate simulation from reality. Synthetic, browser, simulator-only, and hardware claims are distinguished.
- Make rerunning easy. A clean checkout should be enough to reproduce a result.
| Area | Representative work |
|---|---|
| Anomaly detection | oes-resilience, oes512q-latch |
| Evidence infrastructure | spark-evidence-lab, evidence-passport |
| Holography and optics | waveholo, waveholo-telescope |
| RAG reliability | spark-rag-guardrail |
| Numerical contracts | oes32-residual |
| C++ / HLS | oes32-hls, qldpc_decoder_cpp |
| Simulation | qec-scaling-bench, membrane-crystallization |
| Interactive systems | thought-input-game, world-holobridge |
- Results are SYNTHETIC unless a repository names a measured dataset and protocol, in which case they are REPORTED.
- No repository reports a result from quantum hardware, a physical membrane, a medical device, or a field deployment.
- FPGA synthesis, hardware-in-the-loop testing, and independent review are labelled UNRUN where not performed.
- Design goals are labelled TARGET.
- Claim corrections can be opened through the template in sparkainlp-x/.github.
- Applied-AI and research-software roles, remote within Canada or on site in Fredericton
- Research Software Engineer · Applied AI Engineer · ML Evaluation Engineer · Scientific Software Developer
- Collaboration with optics researchers, FPGA/decoder engineers, accessibility-focused partners, and research groups
- Independent replication and focused technical collaborations
Website · LinkedIn · ORCID · Zenodo · Full portfolio
Français (langue maternelle) · English (fluent)
For the complete repository index, release snapshot, project tracks, acknowledgements, and citation records, see PORTFOLIO.md.