I build systems that refuse to answer when they do not know. The products are closed source; the evidence is not.

• VBRL — skill memory for robots. Record how an expert flew; at runtime retrieve similar moments, weight them by how they actually turned out, and issue a command on CPU. Every decision names the stored episodes behind it, before it fires. An empty or distant neighborhood returns unknown and a safe hold. Simulation is live; first hardware is a Crazyflie in a room. Lucerne, CH.
• RLXBT — local infrastructure for reproducible quant research: event-driven backtesting, robustness analysis, and trading RL. Support and security policy.
• AI Price Patterns — pattern research on price series.
• ManusPay — payments.

Quant Research Log is a public record of systematic-trading research: what was tested, what died, and what killed it. Mostly negative results, because most results are negative.

Six of its eleven studies reproduce on a Python interpreter and the standard library alone — no account, no engine, no pip install, no network:

git clone https://github.com/sergio12S/quant-research-log
cd quant-research-log/studies/009-cost-survival
./run.sh

Under a second, and it reprints the study's own numbers.

Reproduced on a clean checkout, covered by a regression that fails without the change, and verified against the project's own suite before submission.

Landed

• initial-d/ml-quant-trading#47 — named the cumulative cost field for what it is, with a backward-compatible alias. Merged.
• bennycode/trading-signals#1316 — Sharpe, Sortino, and maximum drawdown for the performance calculator. Merged.
• QuantConnect/Lean#9716 — a null custom-data source crashed live warmup. The maintainer took the diagnosis and shipped the same guard, extended to three more call sites, as #9718. Merged under their number, not mine.
• initial-d/ml-quant-trading#59 — a protocol-v1 benchmark report, including the case that would not settle across two identical runs.

Open

• microsoft/qlib#2322 — a time-series sampler returned batches mixed across days, because the index labels were reordered and the rows were not. Approved, awaiting merge.
• portfolio-performance/portfolio#5956 — an IB Flex dividend reversal was imported as income. Now reported as a failed item, following the convention the Trade Republic importer already established.
• hummingbot/hummingbot#8427 — DCA backtest stages were all written at the final stage's timestamp.
• bennycode/trading-signals#1315 — a declarative adapter for indicator signals.

Failures that report a plausible number instead of raising. A backtest that returns 12.4% when it should have refused to run is worse than one that crashes, because nothing downstream can tell the difference.

It is the same problem in a drone and in a backtest. VBRL holds position when the neighborhood is empty rather than inventing a maneuver. Study 002 is seven probes for an engine that answers instead of failing. The trading-signals fix replaced a risk ratio of zero with no ratio at all, because zero already means "measured, and there is no edge" — it would have ranked a flawless run below one that lost money.

Confident and wrong is the failure mode worth engineering against.

Based in Lucerne, Switzerland.