Portfolio

Selected engineering proof: what I built, the systems underneath, and the tradeoffs that mattered. Raintree product pages live at raintree.technology.

Agent Safety & Evals

Context & Developer Tools

Data & Intelligence Systems

Product SaaS

Systems & Low-Level

Music & Creative Systems

Policy evals / personal case study

PolicyStrata

deterministic LLM drift tests

What I built

Deterministic regression-testing framework for cross-layer policy drift in LLM data-agent stacks.

Stack

Python 3.10+ / Pydantic / PyYAML / psycopg / pytest / Hypothesis / mypy / Ruff

Runtime

CLI commands run deterministic suites, generated suites, frozen held-out benchmarks, baselines, ablations, artifact reports, scanners, and demos.
Policy oracle stays independent from manifests, grammar, validators, SQL compilers, database controls, and release filters.
Reproduction scripts rebuild paper-facing evidence tables and verify packaged artifacts from a clean checkout.

Data model

Domain fixtures define principals, policies, surfaces, semantic tasks, SQL assets, scanner traces, and benchmark suites.
Runs write `traces.jsonl`, `summary.json`, `metadata.json`, frozen benchmark manifests, and minimized witness JSON.
Scanner examples cover Postgres/dbt traces while deterministic domains cover support SaaS, finance SaaS, and analytics ClickHouse shapes.

Techniques

Cross-layer responsibilities are checked separately so manifest drift, validator drift, compiler drift, database drift, and release drift can be localized.
Failure minimization reduces generated failures into small witnesses that preserve the violated obligation.
Frozen benchmark manifests guard held-out suites from accidental detector or generator drift.
Scanner mode treats PolicyStrata as a release gate over exported traces rather than as an authorization boundary.

Verification

Paper-facing artifact reports 1720/1720 non-clean injected cases with 0 false positives on 80 clean controls.
PyPI package exposes the `policystrata` CLI and typed Python modules.
GitHub release ships SHA256-pinned paper and artifact archives.
Tests cover benchmark freezing, compiler behavior, database assets, detection, policy logic, scanner config, evidence output, and runner behavior.

Portfolio

Selected engineering proof: what I built, the systems underneath, and the tradeoffs that mattered. Raintree product pages live at raintree.technology.

Agent Safety & Evals

Context & Developer Tools

Data & Intelligence Systems

Product SaaS

Systems & Low-Level

Music & Creative Systems

Policy evals / personal case study

PolicyStrata

deterministic LLM drift tests

What I built

Deterministic regression-testing framework for cross-layer policy drift in LLM data-agent stacks.

Stack

Python 3.10+ / Pydantic / PyYAML / psycopg / pytest / Hypothesis / mypy / Ruff

Runtime

CLI commands run deterministic suites, generated suites, frozen held-out benchmarks, baselines, ablations, artifact reports, scanners, and demos.
Policy oracle stays independent from manifests, grammar, validators, SQL compilers, database controls, and release filters.
Reproduction scripts rebuild paper-facing evidence tables and verify packaged artifacts from a clean checkout.

Data model

Domain fixtures define principals, policies, surfaces, semantic tasks, SQL assets, scanner traces, and benchmark suites.
Runs write `traces.jsonl`, `summary.json`, `metadata.json`, frozen benchmark manifests, and minimized witness JSON.
Scanner examples cover Postgres/dbt traces while deterministic domains cover support SaaS, finance SaaS, and analytics ClickHouse shapes.

Techniques

Cross-layer responsibilities are checked separately so manifest drift, validator drift, compiler drift, database drift, and release drift can be localized.
Failure minimization reduces generated failures into small witnesses that preserve the violated obligation.
Frozen benchmark manifests guard held-out suites from accidental detector or generator drift.
Scanner mode treats PolicyStrata as a release gate over exported traces rather than as an authorization boundary.

Verification

Paper-facing artifact reports 1720/1720 non-clean injected cases with 0 false positives on 80 clean controls.
PyPI package exposes the `policystrata` CLI and typed Python modules.
GitHub release ships SHA256-pinned paper and artifact archives.
Tests cover benchmark freezing, compiler behavior, database assets, detection, policy logic, scanner config, evidence output, and runner behavior.

Selected projects

What I built

Stack

Selected projects

What I built

Stack