Systems I've shipped, experiments I've run.

Market text is loud and contradictory; raw sentiment scores rarely survive contact with a real decision.

An NLP pipeline with a human review step on uncertain cases, producing structured, auditable signals.

Cleaner, more trustworthy signals than a fully-automated baseline. [ metrics to come ]

The human-in-the-loop step isn't a crutch — it's where the system earns its trust.

Most agents hide their reasoning. For expert work, that's a non-starter — you can't be accountable for a black box.

Agents that plan, retrieve, and act while exposing every step for inspection and intervention.

In active development — the current main thread of my work.

Autonomy is not the goal; leverage with legibility is.

Clinical models predict, but rarely justify. A recommendation a clinician can't audit is a recommendation they shouldn't take.

Couple prediction with retrieval of grounding evidence, surfacing the rationale and its sources alongside a calibrated confidence.

Auditable explanations that hold up to expert review. [ quantitative results to come ]

The explanation has to fit inside the clinician's time budget, or it doesn't exist.

Grid operators need forecasts, but a point estimate hides the risk that actually drives their decisions.

Deep sequence models that output calibrated distributions, stress-tested with synthetic time-series where real data is thin.

Forecasts that carry honest uncertainty into the decision. [ metrics to come ]

Synthetic data is a modeling assumption — name the belief it encodes, or it will bite you later.