Fig. 1Two lanes: capture and paper trading on the Pi, research offline on public data. Schematic.
A counted search of 1,871 trials, pre-registered in a hash-chained ledger from 12 Oct. The one surviving market-neutral book scored Sharpe 3.45 on the 2026 holdout (backtest, after costs), and my own adversarial review then downgraded it: the paper reports the result and the downgrade.
Rigor
Every result deflated for 1,871 counted trials; from 12 Oct 2026 every test hash-chained into a ledger, with its pass and kill bars, before it ran
Best result
Holdout Sharpe 3.45, +23.0 bp/day net, HAC t 2.61, worst drawdown 7.9% (backtest); downgraded on my own review because part of its design had seen the window
Rust · Redis Streams · Python, on one Raspberry Pi 5
Fig. 2High-level data flow of the platform. Schematic.
Production crypto market-data platform, co-developed in a team of two: I implemented 19 of its 33 Rust exchange integrations and the shared normalization layer.
An LLM contract-matching pipeline over ≈ 255k Kalshi and Polymarket markets, feeding a fee-aware arbitrage scanner. Built solo, with 529 automated tests.
Detector for a pattern consistent with spoofing on a crypto perpetual: over 14 h, about 97% of 4,961 large placements were cancelled essentially unfilled.
Scope
Solo project, built for traders: specification, architecture, analysis direction and deployment
Team entry: I built the fair-value models, quoting logic and research tooling. My team reached Phase 2. Round-1 test-run PnL rose from 1,817 to 4,789 XIRECs.
Scope
Team entry: built the fair-value models and quoting logic; wrote the trading code and research tooling
Four semesters at WUTIS: led a 3-person CRNN volatility team and a 5-person QNN portfolio team, and taught coding workshops; both teams I led won their semester pitches.
Scope
Team member, then project lead of two teams; workshop leader and recruiter
An AI pipeline that turns legacy Natural/Adabas banking code into as-is docs and review-ready Java/OSB integration code: 5 use cases, 10+ connectors delivered.
Scope
Software Engineer
Natural / Adabas · Java · Oracle Service Bus · GitHub Copilot CLI
BSc thesis graded 1: Qiskit and PyTorch QNNs forecasting returns and covariances of five assets, with a pipeline that reproduces the thesis results bit for bit and is benchmarked on 121 held-out days.
Method
Parametrized quantum circuits trained through PyTorch, 32 Optuna trials per target, exact and noisy Aer simulation
Data
Daily bars of BNDX, GLD, IBM, NVDA and USO, split chronologically 359/119/121
An exact Gaussian convolution restored the martingale property that quadrature broke on the real smile (max |z| 17.8 → 1.8) and cut the fixed point from 254/499 to 10/12 iterations.
Data
SPX close of 15 Aug 2019: 1,026 calls and puts, 246/65/67 cleaned OTM quotes for three expiries
Method
Theorem 2.1 fixed point, exact large-step Monte Carlo, hat-bump vega Greeks checked against bump-and-revalue
A fail-closed gate between LLM extraction and the build refuses missing items, dangling edges, cycles and broken LaTeX; the live demo maps all 18 results of my BSc thesis.
Evaluation
69 automated tests (46 pytest + 23 node:test) and a no-LLM fixture pipeline in CI