I don't publish a number I can't re-run.
I build the instruments that measure whether database and AI agent systems actually do what they claim — then publish the experiment alongside the result, so you can reproduce it or show me where I'm wrong.
Portfolio · Writing · LinkedIn · ORCID
sql-benchmarks-dagster · pip install sqlbenchdag · MCP server in the official registry
Every experiment is a capsule addressed by an 8-character SHA-256 fingerprint over its config, SQL, and every line of measurement-relevant Python. Change the method, the ID changes. Cold-cache execution, hard row-count assertions, an integrity seal on every capsule — and OpenTimestamps proofs anchored to Bitcoin on the four published Quack capsules.
| Finding | Capsule |
|---|---|
| DuckDB-over-Quack (pushdown) beats PostgreSQL at every scale past the noise floor — 4.4× @100K, 6.1× @1M, 13.2× @10M rows | 902d1277 |
| Attach-mode overhead grows with scan size (2.6× @100K → 9.5× @10M); pushdown stays flat at ~2× | b8e2bfaf |
| That flat ~2× residual is reduced server-side parallelism, not protocol transport | 25b0e134 |
| Attach mode cannot execute multi-table joins at all; pushdown holds ~1.8× on a 3-way TPC-H join | b198363e |
Full published-capsule index →
I study AI agent systems by instrumenting them and reading the traces, not by asking them how they did.
| bauplan-agent-parse-study | Per-turn agent traces against a commercial lakehouse platform. Surfaced three defects I reported upstream — #369, #370, #371 — including an error message that actively sends an agent in the wrong direction. |
| podcast-rag | Ask a question about a podcast, get an answer and a link to the exact second someone said it. LanceDB, local-first, no cluster and no API key. Built a labelled eval set and found it answering 17 of 20 questions its sources couldn't support; recalibrated precision 0.56 → 0.87. |
| ai-agent-utils | Boilerplate and security guidelines for collaborating safely with autonomous coding agents. |
| awesome-pedantic-medallion | Community-curated playbooks for building shared understanding. |
- Measuring Quack: DuckDB's New Client-Server Protocol — independent measurements of a protocol at beta, every number capsule-backed
- The Blackboard and the Library — agents need two kinds of memory: the blackboard you erase, and the library that persists
- A
READMEis a Set of Falsifiable Claims — if your README says it works, that's a claim, and claims can be tested - Inside
~/.claude/: Structure is Not Security — what a coding agent writes to your disk, in plaintext - The "Blueberry Muffin Exploit" — semantic injection into agent memory, disclosed
- Neo4j NODES 2026 — Structure Is Not Security: Poisoning Graph-Based Agent Memory Through the Extraction Pipeline · Nov 12, 2026
- Data in the D 2026 — Measure What Matters · Detroit, Oct 16–17, 2026
- Canadian Women in Cybersecurity 2026 — Consensus Contagion: Status Bias and the RLHF Tax in Agentic AI Routing Nodes · Toronto, May 26, 2026
- Neo4j NODES 2025 — Is My Brain a Knowledge Graph? · Nov 6, 2025