Joshua Ha
Financial Controller
Bash
C++
MATLAB
JavaScript
Rust
C
Cython
Python
Data Pipeline & Feature Engineering: 5/10
MLOps & Deployment: 4/10
Computational Efficiency: 4/10
Active 1 day ago
Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
A Middle-level Python generalist focused on building real-time edge computer-vision prototypes and local UI integrations. The strongest proven skill is implementing a real-time drowsiness detection pipeline with MediaPipe, Picamera2, model calibration, and local IPC as implemented in detection/detect.py. There is little to no public evidence of production backend patterns such as databases, schema migrations, formal API versioning, or observability/alerting infrastructure.
Technical skills
Bash
C++
MATLAB
JavaScript
Rust
C
Cython• Middle
Python• Middle
MATLAB
Simulink
Cython
Kivy
Kotlin
StateFlow
AI/ML
Claude
Claude Code
Computer Vision
OpenCV
MediaPipe
Scikit-learn
NumPy
Speech Recognition
DevOps
CI/CD
Docker
Git
Robotics
ROS2
ROS
Perception
Sensor Fusion
Simulation
Timeline
University of California (Riverside Campus)
Bachelor's Degree •
Computer Science
Embedded Systems Team Member
•
Middle
UCR Highlander Racing
•
Full-Time
Designed a C-based finite state machine for a high-voltage EV battery management unit to coordinate contactor logic and fault handling. Implemented embedded firmware for an ARM STM32 platform (STM32CubeIDE), managing SPI and UART communication with battery-monitoring ICs. Verified safe operation behavior across dynamic load states through embedded testing with the team.
C
Propulsion Controls and Modeling Team Member
•
Middle
UCR EcoCAR EV Challenge Team
•
Full-Time
Directed an agile team focused on hardware-in-the-loop testing for an EV propulsion controller. Built and verified Simulink/Stateflow models for MIL validation, including automated unit testing tied to coverage goals. Implemented a Git-integrated regression testing framework using Python and MATLAB scripts to support CI/CD for MIL simulations. Supported validation of CAN/CAN FD and Ethernet TCP/IP communications in a QNX-based environment.
Simulink
StateFlow
C++
Python
Git
CI/CD
Middle AI/ML Engineer
Confidence: High ML Engineer
Embedded ML Engineer (senior-level practical) specializing in real-time driver drowsiness detection and UI integration. The strongest proven skill is building production-oriented, low-latency computer vision inference pipelines with per-user calibration and head-pose suppression as implemented in detection/detect.py (feature extraction, calibration, MediaPipe integration and Picamera2 capture). There is little evidence of training infrastructure, experiment tracking, unit tests, or production model lifecycle management in public code.
Model Architecture & Training
2/10
How well models are designed and trained
Minimal model-building evidence; the code loads a pre-trained scikit-learn pipeline and uses it for inference but provides no training loop, hyperparameter tuning, or experiment artifacts.
Evidence
sentinel-drive-assist/detection/detect.py: joblib.load(MODEL_PATH) and pipeline.predict_proba(cal_features)[0][1]
Data Pipeline & Feature Engineering
5/10
How data is prepared for models
Solid feature engineering and calibration for a real-time CV pipeline; explicit EAR/MAR/head-pose features, per-user calibration, and feature normalization are implemented.
Evidence
sentinel-drive-assist/detection/detect.py: build_feature_vector / eye_aspect_ratio / mouth_aspect_ratio
sentinel-drive-assist/detection/detect.py: apply_calibration and build_model_input
Experimentation & Evaluation
1/10
How results are measured and tested
Almost no experimentation or evaluation infrastructure is present; there are informative runtime prints but no tracking, baselines, or reproducible experiment scripts.
Evidence
sentinel-drive-assist/detection/detect.py: prints about threshold and calibration but no W&B/MLflow/experiment artifacts
MLOps & Deployment
4/10
How models are shipped to production
Practical deployment engineering for an embedded system is evident with multi-process orchestration, lightweight IPC, and model serialization, but lacking production-grade serving, versioning, or monitoring.
Evidence
sentinel-drive-assist/src/Sentinel.py: launches detection and voice subprocesses and runs an HTTP server for IPC
sentinel-drive-assist/detection/detect.py: _start_command_server and _send_ui_command for control and alert callbacks
sentinel-drive-assist/detection/detect.py: joblib model serialization usage
Computational Efficiency
4/10
How efficiently computing resources are used
Practical efficiency choices for edge hardware are present - scaled capture resolution, frame skipping, smoothing windows and calibration to reduce false positives and computation.
Evidence
sentinel-drive-assist/detection/detect.py: CAPTURE_WIDTH/CAPTURE_HEIGHT configuration and Picamera2 preview configuration
sentinel-drive-assist/detection/detect.py: FRAME_SKIP, SMOOTHING_WINDOW and ALERT_FRAMES usage
Research Depth & Innovation
2/10
Depth of research and new ideas
Uses established CV techniques (MediaPipe landmarks, solvePnP head pose, EAR/MAR heuristics) but no novel algorithms or paper-level reproduction work.
Evidence
sentinel-drive-assist/detection/detect.py: build_face_landmarker using MediaPipe and get_head_pose using cv2.solvePnP / Rodrigues
Expertise
Computer Vision & Image Analysis• Middle
Audio & Speech Processing• Middle
Industries
Transportation & Logistics• Middle
Technologies
Python• Middle
C++
MATLAB
Cython• Middle
Claude
OpenCV
Claude Code
Scikit-learn
Computer Vision
Speech Recognition
CI/CD
NumPy
Git
Docker
MediaPipe
Kivy
Simulink
Recommendations
- Develop edge-facing computer vision components and integrations for embedded platforms (camera capture, MediaPipe, head-pose heuristics).
- Build real-time voice-activated control and UI glue for safety-critical consumer devices using Vosk/speech engines and lightweight IPC.
- Prototype and validate applied ML features where inference and UX constraints matter, such as in-vehicle alerts and failsafe behaviors.
- Harden the project into production by adding reproducible training scripts, experiment tracking, automated tests and lightweight model versioning.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Backend Developer
Confidence: Medium Generalist
A Middle-level Python generalist focused on building real-time edge computer-vision prototypes and local UI integrations. The strongest proven skill is implementing a real-time drowsiness detection pipeline with MediaPipe, Picamera2, model calibration, and local IPC as implemented in detection/detect.py. There is little to no public evidence of production backend patterns such as databases, schema migrations, formal API versioning, or observability/alerting infrastructure.
API Design
2/10
How well APIs are designed
Minimal API design for local IPC: simple JSON endpoints bound to localhost without versioning, idempotency, or formal error contract; adequate for local control but not designed for backward compatibility or external clients.
Evidence
sentinel-drive-assist/detection/detect.py: _CommandHandler.do_POST handling /set_sensitivity, /set_detection_enabled, /request_reset
sentinel-drive-assist/src/Sentinel.py: _WakeUpHandler.do_POST handling /wake_up, /alert_cleared, /voice_command
sentinel-drive-assist/src/Interface.py: _send_detect_command function posting JSON to http://127.0.0.1:5001
Data Layer & Database
Working with databases
Not evidenced in public code
Scalability & Performance
3/10
Handling load and speed
Local performance considerations are present (frame skipping, smoothing window, limited-resolution capture, non-blocking posts), but there is no evidence of system-level scalability planning, caching strategies, or load-test driven optimization.
Evidence
sentinel-drive-assist/detection/detect.py: FRAME_SKIP, SMOOTHING_WINDOW and use of collections.deque for smoothing/probability buffering
sentinel-drive-assist/detection/detect.py: Picamera2 configuration with CAPTURE_WIDTH/CAPTURE_HEIGHT and time.sleep warmup
sentinel-drive-assist/detection/detect.py and src/Interface.py: non-blocking HTTP posts using threading.Thread to avoid blocking the capture loop
System Architecture
3/10
Overall system structure
Reasonable modular decomposition into detection, UI, orchestrator, and voice processes with explicit IPC via localhost HTTP and subprocessing; clear separation of responsibilities but not a production distributed architecture.
Evidence
sentinel-drive-assist/src/Sentinel.py: starts detect.py and voice_activation.py as subprocesses and runs a local HTTP server for UI control
sentinel-drive-assist/detection/detect.py: separate command HTTP server (_start_command_server) and alert callbacks to UI
sentinel-drive-assist/src/voice_activation.py: standalone process that sends commands to the UI via HTTP, demonstrating deliberate process separation
Security & Auth
2/10
Protecting data and access
Basic security hygiene for a local-only system (binding to 127.0.0.1 and using timeouts) but no authentication, authorization, input schema validation, or secrets handling for networked services.
Evidence
sentinel-drive-assist/detection/detect.py: HTTPServer created with ('127.0.0.1', 5001)
sentinel-drive-assist/src/Sentinel.py: HTTPServer created with ('127.0.0.1', 5000)
sentinel-drive-assist/detection/detect.py and src/Interface.py: urllib.request.urlopen calls include timeout parameters but JSON payloads are broadly accepted without schema validation
Reliability & Observability
3/10
Stability and monitoring
Some reliability patterns exist such as non-blocking posts, process lifecycle management, camera and model cleanup, and timeouts, but there is limited structured logging, metrics, retry/backoff policies, or observability tooling.
Evidence
sentinel-drive-assist/src/Sentinel.py: terminates subprocesses on shutdown and _do_stop/_do_play manage alarm process lifecycle
sentinel-drive-assist/detection/detect.py: face_landmarker.close(), picam2.stop() and handling of headless mode and calibration fallbacks
sentinel-drive-assist/src/voice_activation.py: ACTIVE_TIMEOUT auto-deactivation and handling of queue.Empty in audio loop
Expertise
Python• Middle
Microservices & API Architecture• Middle
Messaging & Real-time• Middle
Industries
Transportation & Logistics• Middle
Recommendations
- Develop embedded real-time computer-vision prototypes and device-side integrations where local IPC and low-latency inference are primary requirements.
- Build standalone voice-activated control systems and Kivy-based user interfaces for hardware prototypes and demos.
- Implement small-scale orchestration for local services, improving robustness with structured logging, retries with exponential backoff, and graceful shutdown handling.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Robotics Engineer
Confidence: Medium Perception Engineer
Perception engineer at a middle level specializing in real-time in-cabin driver monitoring and drowsiness detection using vision and audio integration. The strongest proven skill is building a runnable vision pipeline and UI integration as implemented in detection/detect.py (MediaPipe FaceLandmarker, EAR/MAR calibration, timestamped inference) together with interprocess HTTP control in src/Sentinel.py and src/Interface.py. Public code lacks ROS or DDS middleware, end-to-end latency/jitter measurements, formal control algorithms, and automated test coverage or CI-based HIL simulation evidence.
Motion Control & Kinematics
2/10
Controlling robot movement
Basic kinematics-like work limited to head-pose estimation using solvePnP with explicit singularity handling; no robot kinematics or IK/FK stacks.
Evidence
sentinel-drive-assist/detection/detect.py: get_head_pose - uses cv2.solvePnP and cv2.Rodrigues with sy-based singularity branch
Perception & Sensor Fusion
5/10
Understanding sensor data
Substantial perception work: a complete vision pipeline using MediaPipe face landmarking, calibrated EAR/MAR features, timestamped video inference and smoothing; no advanced probabilistic filters (EKF/UKF) shown.
Evidence
sentinel-drive-assist/detection/detect.py: build_face_landmarker - MediaPipe FaceLandmarker VIDEO mode with timestamped detect_for_video
sentinel-drive-assist/detection/detect.py: build_feature_vector / apply_calibration - EAR/MAR computation and baseline calibration logic
sentinel-drive-assist/detection/detect.py: main - timestamp_ms, prob_buffer smoothing, FRAME_SKIP and calibration loop
ROS & Middleware
2/10
Robot software framework skills
Lightweight middleware patterns implemented via local HTTP servers and non-blocking threads for UI <-> detection IPC; no ROS/DDS/lifecycle or QoS evidence.
Evidence
sentinel-drive-assist/detection/detect.py: _CommandHandler class and _start_command_server for control endpoints on port 5001
sentinel-drive-assist/src/Sentinel.py: _WakeUpHandler and _start_server - HTTP-based callbacks from detection
sentinel-drive-assist/src/Interface.py: _send_detect_command - non-blocking HTTP posts to detection server
Control Algorithms
2/10
Algorithms that steer robots
Application-level control and safety logic present (hysteresis, counters, recovery timers, suppression rules) but no formal control algorithms (PID/MPC/LQR) or tuning traces.
Evidence
sentinel-drive-assist/detection/detect.py: drowsy_count, ALERT_FRAMES, recovery_start and RECOVERY_SECS gating logic
sentinel-drive-assist/detection/detect.py: suppression logic using pitch/yaw thresholds and EAR_CLOSED_OVERRIDE
Real-time Systems
2/10
Precise real-time control
Real-time engineering awareness (frame skipping, timestamped inference, Picamera2 usage, threading) but no deterministic timing measurements, jitter/latency data, real-time executors or ros2_control usage.
Evidence
sentinel-drive-assist/detection/detect.py: FRAME_SKIP, timestamp_ms computed for detect_for_video, use of picamera2 for constant-size capture
sentinel-drive-assist/detection/detect.py and src/Sentinel.py: threading used for HTTP servers and non-blocking posts
Simulation & HIL Testing
3/10
Testing robots in simulation
Includes a standalone cycle-accurate Tomasulo CPU simulator demonstrating simulation architecture and modular pipeline stages, but no robotics HIL or CI-based sim testing.
Evidence
tomasulo-sim/src/sim.py: main simulation loop, FU pipelines, CDB arbitration and termination checks
tomasulo-sim/src/struct.py: clear/new RS/LS/FU factories and tag/issue generators
tomasulo-sim/src/exec.py: dispatch, advance_fus and collect_finished logic
Verified artifacts
Expertise
Computer Vision & Perception• Middle
Industries
Hardware• Middle
Transportation & Logistics• Middle
Technologies
Simulation
Rust
C
ROS
ROS2
Sensor Fusion
Perception
Recommendations
- Add automated tests and benchmark runs that measure end-to-end latency, frame processing time, and jitter on the target hardware (Picamera2 on Raspberry Pi) to quantify real-time behavior.
- Produce a small HIL or dataset collection pipeline (rosbag-like or recorded video/audio logs) and unit/regression tests for the perception model and calibration to support iterative model tuning.
- If the goal is vehicle integration, migrate IPC to a resilient middleware (ROS2 or a QoS-aware DDS) and add lifecycle handling and watchdogs for safety-critical signals.
- Capture and version calibration artifacts (baseline EAR/MAR per user) and add tooling to visualize tuning curves and false-positive/false-negative rates for threshold selection.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: