Independent Research

A CI-native tool for catching LLM inference cost regressions before they reach production.

• Why - LLM changes can pass quality checks while becoming slower, longer, or more expensive.
• Approach - Runs fixed prompt suites, records token/cost/latency metrics, and compares pull-request runs against a saved baseline.
• Outcome - Produces pass/fail gates and reviewable Markdown/JSON reports for cost-per-success, token drift, and latency regressions.
• Methods & checks - Provider adapters, rate cards, deterministic harness settings, repeated trials, practical thresholds, and statistical comparison tests.
• Artifacts - Manuscript under preparation ·  Zenodo Code DOI

Trajectory-based study of whether structural regime changes can be detected without giving the detector access to the governing equations.

• Why - Recorded trajectories are often available before clean equations, Jacobians, or model internals.
• Approach - Built trajectory-derived indicators, combined them into a structural score, and extracted boundaries across parameter sweeps.
• Outcome - On FitzHugh-Nagumo and autonomous forced van der Pol tests, the structural boundary appeared before the qualitative transition.
• Methods & checks - Change-point extraction, lead-distance analysis, nuisance sweeps, robustness tests, ablations, and manifest validation.
• Artifacts - Preprint PDF · Supplement · Zenodo Preprint DOI · Zenodo Code DOI

Benchmark/diagnostic for learned equations of state. Four identity checks (monotonicity, stability, Clapeyron slope, acoustic speed), JSON scoring, and guardrails near critical regions.

• Why - Accuracy metrics can hide violations of basic physical identities.
• Approach - Minimal API; per-check assertions emit structured failures that map directly to the violated identity.
• Outcome - CI-ready tests that block physically impossible predictions before deployment.
• Methods & checks - Finite-difference identities; envelope/phase-boundary sanity tests; critical-region clamping.
• Artifacts - Preprint PDF · Zenodo Documentation DOI · Zenodo Code DOI

Negative-results study showing that seemingly accurate GNNs fail global thermodynamic consistency; proposes a hybrid fallback with classical libraries.

• Why - Small pointwise error ≠ integrable or identity-preserving fields.
• Approach - Stress tests with identity checks; trigger a classical fallback policy when tests fail.
• Outcome - The useful result was not a better GNN; it was a set of failure cases and a fallback rule for when a learned property model should hand off to a classical library.
• Methods & checks - Integrability tests, common-tangent/Maxwell constructions, cross-library comparisons.
• Artifacts - Preprint PDF · arXiv

A 79-study synthesis on energy in autonomy stacks (planning, perception, middleware, HW acceleration) across mobile and manipulator platforms.

• Why - Robotics papers report speed/latency but rarely in comparable energy terms.
• Approach - Systematic screening; taxonomy of techniques (scheduling/DVFS, refactoring, off-loading, task-level policies); reporting checklist.
• Outcome - Quantified gaps (Wh/mission, energy-latency trade-offs, standard workloads) and under-measured components (sensing, comms, OS); full replication package.
• Methods & checks - Inclusion/exclusion rationale; code + data to regenerate figures; foundation for REB-1’s Wh→gCO₂ framing.
• Artifacts - Preprint PDF · arXiv · Zenodo Code DOI

Micro-benchmark for measuring power and carbon footprint of robotics workloads; standardized 60-second traces and tidy logs for repeatable comparison.

• Why - Teams lack a consistent yardstick for energy across perception/control stacks.
• Approach - Fixed-duration runs; synchronized power + CPU/GPU counters; ready-to-plot summaries.
• Outcome - Repeatable numbers that enable energy-aware choices and rank stability across hardware.
• Methods & checks - Normalized telemetry; per-board calibration notes; Wh → gCO₂ conversion.
• Artifacts - Zenodo Code DOI