I architect production-grade systems at the intersection of low-level hardware control and state-of-the-art machine learning. My expertise spans from optimizing C++ physics engines with SIMD instructions to deploying Transformer-based detection models on edge devices.
|
|
C++17 |
Python |
PyTorch |
CMake |
A research-grade quadrotor simulation platform engineered for training autonomous flight controllers via deep reinforcement learning. The architecture couples a high-performance C++ physics core with a modular Python training pipeline.
Physics Engine Architecture (C++)
| Module | Implementation |
|---|---|
| Rigid Body Dynamics | Full 6-DOF Newton-Euler equations with quaternion normalization |
| Numerical Integration | RK4, RK45 adaptive step-size, Velocity Verlet with sub-stepping |
| Rotor Aerodynamics | Blade Element Theory coupled with momentum theory |
| Ground Effect | Height-dependent thrust augmentation model |
| Motor Dynamics | First-order lag with ESC response, thermal derating curves |
| Battery Model | State-of-charge curves, internal resistance, voltage sag |
| Wind Turbulence | MIL-SPEC Dryden model with configurable intensity scales |
class alignas(64) Quadrotor {
void stepWithSubStepping(const MotorCommand& command, double agentDt);
void stepAdaptive(const MotorCommand& command, double& dt);
Vec3 computeTotalThrust() const noexcept;
Vec3 computeAerodynamicForces(const Vec3& velocityBody) const noexcept;
Vec3 computeGyroscopicTorque() const noexcept;
Mat3 computeInertiaInverse() const noexcept;
};Training Pipeline Architecture (Python)
| Component | Description |
|---|---|
| TD3 Agent | Twin Delayed DDPG with target policy smoothing and delayed updates |
| Experience Replay | Sum-tree based prioritized sampling with importance weighting |
| Environment | Gymnasium-compliant with configurable task modes |
| Domain Randomization | Stochastic variation of mass, inertia, motor constants |
| Curriculum Learning | Progressive difficulty scaling for stable policy convergence |
| Vectorization | Parallel environment execution via SubprocVecEnv |
class TD3Agent:
def select_action(self, state: np.ndarray, exploration: bool = True) -> np.ndarray:
with torch.no_grad():
action = self.actor(state)
if exploration:
action += self.exploration_noise.sample()
return action.clamp(-1.0, 1.0).cpu().numpy()
|
Python |
PyTorch |
OpenCV |
Streamlit |
An end-to-end pipeline for automated pothole detection, metric depth estimation, and severity classification from monocular imagery. Designed for municipal infrastructure assessment with priority-based maintenance scheduling.
Detection and Depth Pipeline
| Stage | Technology |
|---|---|
| Object Detection | YOLOv8 fine-tuned for multi-class road damage |
| Depth Estimation | Depth Anything v2 for dense monocular depth |
| Plane Fitting | RANSAC-based ground plane estimation |
| Depth Refinement | Guided filtering for edge-preserving smoothing |
| Metric Calibration | Camera intrinsics for physical measurements |
class PotholeAnalyzer:
def analyze_image(self, image_np_bgr: np.ndarray) -> List[Dict]:
detections = self._detect_potholes(image_rgb)
depth_map = self._get_depth_map(image_rgb)
ground_plane = self._fit_ground_plane(depth_map)
return self._compute_severity_metrics(detections, depth_map, ground_plane)Severity Classification System
| Feature | Method |
|---|---|
| Dimensional Analysis | Metric extraction of length, width, depth |
| Confidence Mapping | Detection probability weighting |
| Multi-Factor Scoring | Geometric + contextual features fusion |
| Priority Ranking | Automated maintenance queue generation |
|
Python |
PyTorch |
OpenCV |
NumPy |
A low-latency ball tracking system for sports analytics combining Transformer-based detection with probabilistic state estimation for robust tracking under occlusion, motion blur, and rapid direction changes.
Detection Architecture
| Component | Implementation |
|---|---|
| Backbone | RF-DETR (Region-Focused Detection Transformer) |
| Multi-Scale | Adaptive resolution for varying object sizes |
| Precision | FP16 inference with TensorRT optimization |
| Temporal Fusion | Detection history for confidence smoothing |
| ROI Processing | Dynamic region-of-interest for computational efficiency |
Tracking Architecture
| Component | Implementation |
|---|---|
| State Estimator | Extended Kalman Filter with 6-state model |
| Gating | Mahalanobis distance for statistical outlier rejection |
| Noise Adaptation | Dynamic process/measurement noise tuning |
| Smoothing | One-Euro Filter with adaptive cutoff frequency |
| Hypothesis Management | Multi-hypothesis tracking for ambiguity resolution |
class ExtendedKalmanFilter:
def predict(self, dt: float) -> np.ndarray:
self._update_transition_matrix(dt)
self.x = self.F @ self.x
self.P = self.F @ self.P @ self.F.T + self.Q
return self.x[:2]
def update(self, z: np.ndarray, confidence: float) -> np.ndarray:
d = mahalanobis_distance(z, self.x[:2], self.P[:2, :2])
if d < self.gate_threshold:
self._kalman_update(z, confidence)
return self.x[:2]
| Domain | Technologies | Applications |
|---|---|---|
| Physics Simulation | C++17, SIMD, RK4/RK45, Quaternions | Quadrotor dynamics, rigid body mechanics |
| Reinforcement Learning | TD3, DDPG, PER, Curriculum Learning | Autonomous control, policy optimization |
| Object Detection | RF-DETR, YOLOv8, DETR, Vision Transformers | Sports analytics, infrastructure inspection |
| State Estimation | EKF, Mahalanobis Gating, One-Euro Filter | Multi-object tracking, sensor fusion |
| Depth Estimation | Monocular depth, RANSAC, Guided Filtering | 3D reconstruction, metric measurement |
| Edge Deployment | TensorRT, ONNX, FP16/INT8, Raspberry Pi | Real-time inference, embedded AI |
Building intelligent systems that bridge physical reality and computational intelligence.
El Salvador • Systems Engineer • Open to Collaboration