Data engineer. I build data platforms on AWS and, outside work hours, small autonomous systems and research tooling — mostly Python, always with a bias for things that can be checked rather than believed.
The rule I build to: if a tool makes a claim, it should be able to prove it. Each repo below ships the tests, an evidence file, and the failure modes it does not hide.
factorproof — factor research that refuses to promote noise
Purged walk-forward cross-validation with embargo, IC decay curves, Benjamini-Hochberg FDR across the whole factor family, a deflated Sharpe ratio reported per factor (computed, not gated) — and a promotion gate a factor must satisfy before it is allowed into a portfolio.
$ factor-lab list
Name Category Description
mom_20 momentum 20-day log-return (short-term momentum)
rev_5 reversal Negated 5-day log-return (short-term reversal)
vol_20 volatility 20-day realised volatility, annualised
rsi_14 oscillator 14-day Relative Strength Index (Wilder smoothing)replayproof — offline regression testing for LLM agents
Record an agent run once, re-run it forever with zero API keys. Assert tool-call contracts, gate the token cost of a change, and diff behaviour when you swap the model underneath.
$ agenteval replay --run examples/recordings/sample_run.jsonl --mode strict
$ agenteval drift --baseline base_result.json --current new_result.jsonfitsproof — local inference that proves it fits
A calibrated cost model and an enforced memory budget in front of a from-scratch NumPy inference engine (KV cache, quantisation). When the model does not fit, it says so loudly instead of OOM-ing silently at token 400.
$ fitsproof probe
Probing machine...
bandwidth: 6.69 GB/s
gemm: 314.31 GFLOPS
RAM: 33.55 GB
VRAM: 0.00 GBAn MCP server that gives an AI agent read access to a warehouse through a SQL-AST-enforced column policy, row filters, query budgets, redaction and an append-only audit trail. In progress.
Merged work in other people's repos. 17 PRs in total; these are the ones worth reading.
- aws/aws-cdk#37815 —
fix(glue-alpha): restorenotifyDelayAfterto PySpark and Scala Spark ETL jobs (13k★) - google/zx#1464 —
fix(deps): skipnode:protocol imports in--install(46k★) - teng-lin/notebooklm-py#2262 —
feat(mcp):source_deleteaccepts multiple sources (20k★) - strands-agents/harness-sdk#4067, #4059, #3955 — dropped content blocks, context offloader, llama.cpp usage chunk (9.2k★)
- dlt-hub/dlt#4283, #3928 — table-prefix layout, decimal coercion (6.1k★)
- Cygnux-Labs/Tracekit#474, exmergo/dex#775, ferrumio/pynox#40
Some of these went through review: dex#775 got changes requested, then approved, and
harness-sdk#4067 was approved by a maintainer.
Each repo went through the same iteration protocol — research → implement → evaluate →
adversarial review → mutation testing — with the evidence committed next to the code
(EVIDENCE.md, docs/ADVERSARIAL_REVIEW.md, reports/).
CI runs each suite on Python 3.11 and 3.12, plus ruff check and ruff format --check.
The badges above are live status, not a number I typed and forgot to update.
Where something is unfinished or wrong, the README says so. A repo that hides its gaps is a demo; one that states them is a piece of work.
- AWS data lake engineering — Glue, Athena, S3, orchestration; cost and correctness at scale
- I keep forks of dbt, Dagster and Ibis for reading. They are not contributions.



