Robotics engineer working at the intersection of first-principles physics, numerical methods, and modern ML. I like building systems that are both mathematically rigorous and deployable on real hardware.
Most of what's below follows the same spine: derive the physics → learn what the physics can't close → benchmark it honestly against the classical baseline → port it to something that runs in a control loop. The interesting part is usually where the model loses, so that's what I try to publish.
vehicle-dynamics-estimation · Python / C++ / PyTorch Physics-informed parameter estimation for vehicle braking dynamics. Five estimators on the same data — batch optimiser, EKF, MLP, and two PINNs — with an honest model-mismatch study showing which method wins when. Ships an allocation-free C++ edge port: 3,414× EKF speedup over the Python implementation, which is 99.97 % per-call dispatch — 62 KB binary, zero external dependencies, zero runtime allocations, Python↔C++ parity to 6.7e-9.
engine-map-pinn · Python / PyTorch From a dyno pull to a Nürburgring lap: a PINN reconstructs the full engine torque map from sparse sweeps, then a quasi-steady-state lap sim runs it on real circuit geometry (Silverstone, Spa, Nordschleife). Two findings worth the build — doubling the engine's power buys only 5–6 % of lap time because most of a lap is grip-limited, while re-optimizing the racing line saves 53 s on the Nordschleife. Validation brackets the SVJ's real 6:44.97 record rather than tuning to hit it.
singularity-robust-control · Python / NumPy
Detect an approaching kinematic singularity in a 3R manipulator and switch to a damped, singularity-robust velocity law before joint velocities blow up. The testbed is chosen so the singularity is exact and hand-derivable — det(J) = l₁·l₂·sin(q₂) — and the workspace manipulability map is validated against a closed form via Heron's formula, so the figures can't quietly drift from the algebra.
monocular-vo · Python / PyTorch / OpenCV Monocular visual odometry with metric-scale trajectory recovery from a single calibrated webcam, using Depth Anything v2 depth + ORB matches + PnP-RANSAC — bypassing the scale-ambiguity wall that classical essential-matrix VO hits. 12.96 % scale error on a tape-measured 5 m hallway walk; a pose-graph back-end with loop closure cuts ATE by 36 % on TUM RGB-D.
vision_demos · Python / OpenCV / MLX
Realtime CV on Apple Silicon, building to a closed perception → decision → actuation loop. gesture_bot takes webcam gestures through a debounced state machine (confidence gate, stability requirement, dead-man timeout) to (v, ω) — the same pair as geometry_msgs/Twist — behind pluggable sim / Arduino / HID backends.
aerospace-surrogate · Python / Rust ML surrogates for airfoil aerodynamics, benchmarked against the 100-year-old thin-airfoil baseline — which wins in the linear regime and only collapses (R² = −1.77) through stall, where the surrogate holds R² = 0.77. A nonlinear lifting-line solver lifts 2D polars to 3D finite wings, reproducing the elliptic-wing identity to machine precision. The Rust port runs 9.8× faster than Python and agrees to 1e-10 across 28 parity tests.
synfuel-control · Python / C++ / ESP32 End-to-end sense → model → control → deploy for a thermal-runaway-prone Fischer–Tropsch reactor. Sobol analysis identifies pressure as the dominant runaway driver; a PINN surrogate (R² ≈ 0.996) drives an RTO + PI controller that holds 296 °C through a cooling failure that otherwise runs away to 328 °C. Deployed as a dependency-free C++ engine at 0.98 µs/inference — 4.8× faster than the ONNX Runtime Python path — closed in software-in-the-loop through a modeled analog front-end and an ESP32 node.
flux — interactive 2D/3D visualizations of the math I keep reaching for: Jacobians & manipulability ellipsoids, Kalman filtering, SE(3) screw motion, Fourier, optimization, eigenvectors. Vanilla JS, no build step. wildfire — operational two-stage US West Coast fire prediction: XGBoost ignition (ROC-AUC 0.785, evaluated fully out-of-time) + U-Net spread matching the Google Research NDWS benchmark. Live dashboard. neutrino-ml — ML on real ATLAS Open Data (13 TeV W→ℓν): RDataFrame selection and neutrino-p_z reconstruction, recovering the W mass at 80.38 GeV. quantview — Bloomberg-style quant terminal in the shell. Black-Scholes Greeks, CRR/trinomial lattices, Monte Carlo (GBM + Heston), Markowitz. Pure NumPy/SciPy math layer cross-checked against QuantLib.
Python · C++ · Rust · PyTorch · NumPy / SciPy · Kalman filtering · numerical optimization · physics-informed ML · embedded & edge deployment · CMake · GitHub Actions