Computer Science and Software Engineering graduate, Maynooth University, 2026. Based in Dublin, available now. Graduate software engineering and quantitative developer roles, Dublin or remote across the EU and UK.

Portfolio · CV · LinkedIn · Email

Fixes shipped into other teams' codebases, reviewed by their maintainers and held to their test suites. Each link is the pull request itself.

• hflow (Hebbian Robotics) #288: a dedicated exception and failure kind so transform content refusals stop being retried as infrastructure errors.
• narwhals #3856: the pandas-like backend ignored the schema argument's column selection and order in from_dicts; fixed with six new tests, each proven failing first.
• exchange_calendars #593: the Dubai Financial Market (XDFM) calendar from scratch, sourced from DFM and Nasdaq Dubai circulars and bounded at the 2022 UAE work-week switch.
• holidays #3748: Russian localisation for Latvia's holiday calendar, native-speaker translation.
• yfinance #2936: traced a suspected timezone regression to an intentional, changelogged change, then fixed the real gap in the download() docstring.

Nine more fixes are open under maintainer review across scipy, splink, uvicorn, duckdb-python, quantstats, and soccerdata.

jobq · Go, standard library only Durable job queue from scratch: at-least-once delivery, actor-per-topic concurrency, leases, retries, dead-letter queues, and a CRC-checked, group-committed write-ahead log (550µs to 65µs per op under 16 producers). A crash harness kills the process cold mid-write across five rounds: zero acknowledged jobs lost, zero resurrected. Conservation and unique-settlement invariants hold under the race detector.

Options pricing engine · live explorer Black-Scholes, CRR binomial (European and American) and seeded Monte Carlo pricers cross-validated through seven machine-checked numerical gates (put-call parity to 2.84e-14). Inverts a real SPY option chain to implied volatility and recovers the smile. The explorer re-runs all seven gates on every page load.

Jarvis · Electron, whisper.cpp, PowerShell, Claude agent Desktop assistant with fully on-device speech-to-text and a scheduled, unattended agent pipeline that delivers a daily 08:30 Telegram briefing. Two-way commands over a closed whitelist locked to one chat ID, failing closed on an unknown sender. CI green on windows-latest, gitleaks over the full history.

registry-resolve · Python, Splink, DuckDB Entity resolution across three real Irish open-data registers (821k CRO companies, the Register of Charities, public procurement awards). Deterministic ID joins first, probabilistic matching for the residual, evaluated against 330 blind human-labelled pairs. 224 tests.

irish-property-price-index · dbt, DuckDB · live dashboard Mix-adjusted stratified-median price index over all 797,774 sales in the Property Price Register, 14 dbt models and 91 dbt tests, validated against the CSO's official index. Finding: the raw median overstates national growth by 2.22 index points and understates Dublin's by 3.91.

The hypothesis and the pass/fail gate are written and committed first, then the analysis runs. Five of these have a verdict; all five are published as they came out, and no threshold was widened once the result went the wrong way.

fyp-strategy-engine, verdict DISPROVEN Six-phase test of a strategy whose headline backtest was +$28,400. The configuration is hash-frozen and the runner refuses to start on a mismatch. 17 walk-forward folds per instrument across ~10 years of CME index futures: pooled profit factor 0.905 over 1,402 out-of-sample trades, 90% upper bound below breakeven. 176 tests in CI. The gates also caught a 60-minute timestamp bug in the source dataset.

alpha-signal-lab, verdict NULL · dashboard LightGBM versus plain momentum on 5-day cross-sectional equity returns under purged walk-forward CV and eight automated leakage audits. A deliberately leaky twin shows the gap: honest rank IC 0.0100 versus 0.1506. The honest model does not beat momentum; a SHAP diagnostic explains why.

prompt-placebo, verdict PLACEBO Pre-registered, paired-delta audit of six prompting techniques across 1,438 questions and 23,008 API requests. 0 of 39 comparisons still work on 2026 reasoning models; the only two effects to survive Holm-Bonferroni correction are harms.

football-trajectory, verdict NOT PROVEN · live explorer Monte Carlo career projection over 24,057 FBref player-seasons. Real skill at every horizon, but the 80% intervals cover ~90%, so it fails its own calibration gate by being under-confident and ships labelled that way. 95 tests.

polymarket-favourite-bias, verdict NOT PROVEN 2,418 resolved markets: favourites win 90.6% of the time and still return −0.98% after fees, with a bootstrap interval spanning breakeven. The gate was the repository's first commit.

dublin-bikes-forecast, verdict pending, paused Twice-daily P(bike)/P(dock) forecasts for ~115 stations, each git-committed before its target time and scored against pre-registered baselines. Paused since 23 July 2026 when the cloud host suspended the VM; the ledger shows the gap rather than restarting the clock.

Education: B.Sc. (Hons) Computer Science and Software Engineering, Maynooth University, 2026, 2:1 Honours. Experience: Quantitative Researcher and Software Engineer, DLT Capital, February to July 2025 (~$15K live trading profit from systematic strategies).

I use Claude Code as a pairing partner; the co-author trailers in these histories say so. Every method choice, architecture call, and verdict is mine, and I can walk through the reasoning on any commit.