THREEDEXENGINEERING

THREEDEX / APPLIED INTELLIGENCE

Applied AI.
Engineered
to work.

From a promising model to a working system.
AI agents, computer vision, and model training—with the software, evaluation, and controls around them.

Discuss your AI use case Explore the systems

Independent engineering. Led by Yehor Horai.

01 Agentic systems02 Computer vision03 ML & model trainingOne connected delivery loop

01 / SYSTEMS IN PRACTICE

Under the hood.
Out in the real world.

Explore how the pieces fit together.
Interactive demonstrations and a recorded vision example.

agent-workbench / system mapINTERACTIVE DEMO
CLIENT OUTREACH / PITCHAPP

Prospect intelligence to reviewed draft

Ready
EXECUTION TRACE00/07
    HUMAN GATE

    The flow pauses before any message can enter a delivery queue.

    Architecture-based simulation. No email is sent.

    TENDER RADAR / TED EVIDENCE

    Official notice to human decision

    Ready
    EXECUTION TRACE00/07
      HUMAN GATE

      The flow pauses before a proposal draft or workflow decision.

      Architecture-based simulation. No tender is submitted.

      AGENTIC APPLICATIONS

      Two systems. Visible control points.

      Explore a client outreach pipeline grounded in PitchApp and an evidence-first tender workflow grounded in Tender Radar. Each map separates automated processing from the decision a person must make.

      • Python
      • React
      • SQLite
      • Evidence gates
      vision-lab / basketballRECORDED EXAMPLE
      COMPUTER VISION / TRACKING
      Recorded output · no live inference
      DETECT→STABILIZE→TRACK→VFX
      COMPUTER VISION

      See the signal.
      Handle the noise.

      Custom basketball detection with temporal constraints, impossible-motion rejection, and gap handling. Track positions and masks become inputs for real-time visual effects.

      Open recording
      • YOLO
      • OpenCV
      • PyTorch
      • TouchDesigner
      training-lab / basketball detectorACTUAL TRAINING LOG
      RECORDED RUNRecorded

      YOLO26l · 1280px · batch 6 · 82 completed / 160 planned epochs

      EPOCH57 / 82
      TRAIN BOX—
      VALIDATION BOX—
      BOX LOSS / EPOCHTrain box Validation box
      Basketball detector training and validation box lossRecorded YOLO26l v2 training values through epoch 57 of 82.2.251.130.0014182BEST 57
      SOURCE: RESULTS.CSV
      ML & MODEL TRAINING

      Train with intent.
      Evaluate with context.

      This graph reproduces the 82 completed epochs from the YOLO26l v2 basketball detector run. Switch between loss and mAP, inspect individual epochs, and see the best recorded mAP50–95 at epoch 57.

      • YOLO26l
      • PyTorch
      • CUDA
      • Evaluation

      ALSO IN THE TOOLBOX

      Sign language recognition

      Landmarks, sequence windows, candidate ranking, and controlled prediction triggers.

      PPE & helmet detection

      Custom YOLO training and video inference with thresholds and visual verification.

      02 / ENGINEERING CAPABILITIES

      Build the whole loop.

      Models are a starting point.
      Delivery includes everything around them.

      01

      Agents &
      applied AI

      Turn a business workflow into a grounded, stateful application with deliberate control over what the model can do.

      • RAG & structured model outputs
      • Tool permissions & approval gates
      • Persistent state & failure recovery
      • Scheduling, rate limits & audit trails
      02

      Vision &
      video systems

      Make camera and video signals useful through preprocessing, model inference, temporal logic, and practical UX.

      • Detection, segmentation & tracking
      • Landmarks & sequence classification
      • Confidence & motion constraints
      • Mask export & real-time integration
      03

      Training &
      inference

      Prepare the data, establish an evaluation loop, then package the model into an application that can be operated.

      • Custom training & fine-tuning
      • Dataset splits & failure analysis
      • GPU inference & batch execution
      • Containerized APIs & deployment
      CONNECTED BY SOFTWARE

      Backend APIs, state, interfaces, and inference infrastructure. Python · Docker · SQLite · RunPod · GCP/GKE.

      03 / FROM QUESTION TO SYSTEM

      A clear path through
      an uncertain problem.

      1. 01 /

        Define the behavior

        Inputs, expected outputs, integrations, and the cost of getting it wrong.

        SCOPE & SUCCESS CRITERIA
      2. 02 /

        Establish a baseline

        Inspect the data, separate validation sets, and test a bounded first approach.

        FEASIBILITY & EVALUATION
      3. 03 /

        Build the loop

        Connect models, state, tools, and UI. Make failures and review points visible.

        WORKING PILOT
      4. 04 /

        Make it operable

        Package inference, verify deployment constraints, document, and hand over.

        DELIVERY & HANDOVER

      04 / THE ENGINEER BEHIND THE SYSTEMS

      Hi, I’m Yehor.
      I connect the dots.

      I work across model behavior and software delivery: data, training, evaluation, inference, state, and the interface people actually use.

      My background in Unity and Unreal real-time applications carries into computer vision and AI systems—where camera input, timing, and feedback loops matter as much as the model.

      Computer Science · National Aerospace University, KhAI
      NVIDIA training · LLM applications, real-time video AI, and Jetson

      Let’s talk about your system
      ENGINEERING STACKTOOLS, WITH PURPOSE
      Intelligence
      PyTorch / YOLO / RAG
      MediaPipe / OpenCV
      Orchestration
      Python / SQLite / n8n
      ComfyUI / custom nodes
      Delivery
      Docker / Cog / CUDA
      RunPod / GCP / GKE

      The right tools follow the problem.

      BEFORE WE BUILD

      A few useful
      starting points.

      Do we need a custom model?

      Not necessarily. Start with the required behavior and a baseline using an existing model. Custom training or fine-tuning makes sense when the evaluation and available data justify it.

      What do you need to assess a project?

      A short description of the problem, representative inputs, expected outputs, current software, and deployment constraints. We can then define a bounded feasibility assessment or pilot.

      Can you integrate into our existing software?

      Yes. APIs, persistent state, interfaces, and containerized inference can be scoped alongside model work. Access, hosting, review points, and handover are agreed around your environment.

      LET’S BUILD SOMETHING USEFUL

      Bring the problem.
      Let’s engineer the system.

      Discuss your AI use case

      Start with your data, the behavior you need, and where it needs to run.