CyberPodolia

Digital Desktop

About Me

CyberPodolia

CV/ML Engineer • Python Dev

Senior CG / procedural background (real-time/ Houdini/ Unreal Engine). transitioning into applied CV and edge ML, with a focus on synthetic data, deployment, and robust pipelines.

Focus: synthetic data, computer vision, reliable pipelines, and clean ops (Linux, Docker, Nginx).

Building systems that feel like products.

Links
Video Introduction
Tech Stack
Python FastAPI Flask Typer Playwright pytest SQLAlchemy SQLite PostgreSQL Docker Nginx Gunicorn Linux Git GitHub Actions JavaScript TypeScript Unreal Engine Houdini Blender OpenCV PyTorch ONNX
Webhook Relay API

FastAPI webhook intake + event store, with idempotency and optional relay.

Features

  • HMAC signature verification
  • Idempotency-Key support
  • Relay allowlist + SSRF checks
  • Performance metrics via k6 /perf and Prometheus /metrics

Stack

FastAPI SQLAlchemy Docker pytest
GitHub
QA Automation Suite

Playwright + pytest UI tests with per-test artifacts and HTML reports. Designed to stay readable and stable.

Features

  • Trace/video/screenshot on failure
  • HTML report + attachments
  • Page Objects + stable selectors

Stack

Playwright pytest Python
GitHub
Asset Scan CLI

Asset tree scanner: naming rules, duplicates, JSON/NDJSON reports, metrics.

Features

  • Clear exit codes for automation
  • Duplicate grouping + sorting
  • JSON/NDJSON output
  • Prometheus textfile metrics

Stack

Python Typer pytest
GitHub
Live Contact
About This Site

A Flask portfolio that behaves like a desktop: draggable/resizable windows, layout persistence, and a Telegram-backed contact flow.

What it is

  • Drag & resize widgets
  • Auto-save layout to localStorage
  • Contact form with Telegram API
  • Reset & Share buttons
  • WebGL animated background

Architecture

Flask Gunicorn Nginx Vanilla JS WebGL Telegram
GitHub
Local Planner
Click to enlarge

Offline-first daily planner for Windows: local API + scheduler + tray UI, with SQLite as the source of truth.

Key features

  • Scheduler + alerts: reminders and start/deadline notifications (optional Telegram)
  • KPI dashboard: last N days status mix + time totals
  • Board + task detail pages for daily flow
  • CLI + local API
  • LLM layer for planning/rewrite

Project status

Stage 1 - Personal Prototype: proving the workflow and metrics in a real daily loop.

Tech

Python FastAPI SQLite Telegram Windows
GitHub
Jetson YOLO Realtime: Detection + Tracking
Before optimization
After optimization

Real-time object detection on Jetson Nano 4GB from a USB camera. After optimization, benchmark runs reached up to 28-29 FPS.

Features

  • Custom dataset creation and YOLO training on PC
  • YOLO -> ONNX -> TensorRT FP16 deployment pipeline on Jetson Nano
  • Real-time USB camera inference on Jetson Nano
  • Detection + tracking modes for steadier object trajectories
  • Live preview with bounding boxes, FPS, and latency overlay
  • Headless benchmark mode with JSON metrics and per-stage timings
  • Edge optimization focused on Jetson performance and efficiency

Tech

Python YOLOv8 ONNX TensorRT FP16 Jetson Nano OpenCV
GitHub
AI Orchestration UI (R&D • MVP in progress)

Local-first orchestration UI for reproducible AI runs: deterministic execution, approval gates, traceable artifacts (logs/reports), and KB-aware routing.

Personal R&D on applying scaling through standardization and variability control.

Highlights

  • Deterministic workflow graph with optional approval steps
  • Standard run flow (plan -> implement -> test -> review -> report)
  • Per-run artifacts: logs, reports, outputs
  • Knowledge base integration (RAG-ready)
  • Local-first by default; cloud optional
  • Node-based graph UI (in progress)

Status

Working MVP; active development / expansion.

Stack

Python Desktop UI Workflow Orchestration KB/RAG Deterministic Graph Observability OpenGL
GitHub
SmallSR

Graduation project from the Robot Dreams course. Validation of the idea of CNNs as universal function approximators.

Lightweight teacher-student pipeline for real-time image restoration and super-resolution on low-res camera streams.

Core Features

  • Super-resolution to 1280x960 from 640x480 input (2x, 4:3 preserved)
  • Teacher pipeline with 4x internal upscale + downscale for detail stabilization
  • Denoise and deblur teacher stages for pseudo-ground-truth generation
  • Lightweight student model for fast real-time inference
  • ONNX + TensorRT export/inference support
  • Batch image processing via python/image_pipeline scripts

Stack

PyTorch ONNX TensorRT Python OpenCV
GitHub
Realtime FaceFX

Research project focused on real-time face-processing pipelines, ROI-first runtime design, and integrating native C++ hot paths into a Python application. The current runtime_cuda variant is validated at about 30 FPS in the 720x540 live preset.

Focus

  • Real-time face-processing experiments with ROI-first runtime design
  • Native C++ hot-path integration inside a Python runtime pipeline
  • Triangle warp, color, shading, and composite stages tuned for live performance
  • Profiling overlay and debug passes for pipeline inspection
  • Validated current preset at about 30 FPS in 720x540 live mode

Tech

Python C++ OpenCV MediaPipe Realtime
GitHub
Photogrammetry Specular Cleaner
Click to enlarge

Single-image specular highlight suppression for photogrammetry: reduces glare to improve feature matching and texture consistency.

What it is

  • Batch inference for single images or folders
  • Suppresses specular highlights (shine/glare)
  • Built on a pretrained model from CXH-Research/DHAN-SHR
  • Optional GPU acceleration via CUDA-enabled PyTorch
  • Memory-safe tiled processing for high-res datasets
  • Works well as a fallback when cross-polarized capture isn't possible

Capture note (recommended)

Cross-polarization (parallel/cross frames) is the most reliable way to remove glare; this tool is the single-image option when you can't control lighting.

Architecture

Python PyTorch TorchVision Pillow NumPy Einops
GitHub