DocsProductSolution
Product

Solution

What ResonTech is, why it exists, and how training and inference work in four steps each.

What is ResonTech?

ResonTech is an end-to-end ML ecosystem — a SaaS platform for running training and inference at scale. Three GPU pool types. One platform. Zero infrastructure overhead.

You write model code. Submit a job. Get your model back. Every layer in between — provisioning, distributed execution, fault recovery, artifact delivery — is handled by the platform kernel.

LayerWhat ResonTech handles
Compute provisioningSelects optimal GPU nodes by type, locality, and current load
Distributed executionSplits jobs across nodes, configures NCCL, syncs gradients
Data deliveryWorkers pull training data and scripts from your S3 bucket at job start
Fault recoveryDetects failures in seconds, reschedules, resumes from checkpoint
Artifact returnWrites model weights and logs back to your workspace folder
Inference servingDeploys endpoints with autoscaling, health checks, load balancing

Mission

AI teams waste enormous time managing infrastructure instead of doing science — configuring Kubernetes clusters, debugging NCCL errors, recovering from spot instance preemptions, reconciling GPU billing. Hours that should go to model architecture and iteration.

The mission is to eliminate that entirely.Submit a job. Get your model back. That's it.

i
ResonTech was founded at ETHKyiv 2025, where the team won the DeAI and State Channels tracks with a prototype decentralized GPU marketplace. The core insight: the hardest part wasn't the payment layer — it was orchestration. Reliable, fast, automatic distributed job execution was unsolved for most ML teams.

Training — How It Works

Training runs through four kernel-managed steps. Your code runs unchanged inside this pipeline.

StepWhat happens
01 · UPLOADUpload your training data and scripts to your S3 bucket via the Files page or rclone. Workers pull from your bucket at job start.
02 · WRITEWrite your training scripts. PyTorch, TensorFlow, HuggingFace Trainer — any framework runs as-is. No SDK wrapping.
03 · EXECSubmit via the platform or API. The kernel provisions nodes, configures distributed execution, and starts your job.
04 · RETURNArtifacts and logs are written back to your workspace folder. If a node fails, the job reschedules and resumes from the last checkpoint automatically.
i
Provisioning, worker selection, distributed execution, fault recovery, and artifact delivery are all kernel-managed. No servers to configure. No CUDA setup. No idle GPU bills between runs.

Inference — How It Works

Inference endpoints go live in four steps. Cold start under 10 seconds.

StepWhat happens
01 · LOADPush a checkpoint from your workspace or point at a HuggingFace repo. The model is fetched and warmed on the cluster.
02 · BINDSet GPU type, concurrency limit, and autoscale rules in one config.
03 · LISTENYour endpoint is live. REST API, OpenAI-compatible for LLMs. Cold start under 10 seconds.
04 · SCALEReplicas spin up on demand and scale to zero when idle. You pay only for active inference time.
i
GPU provisioning, load balancing, replica autoscaling, health checks, and rolling deployments are all kernel-managed. Your endpoint is a URL, not an infrastructure project.

Three Cluster Types

Every workload has a matching pool. Choose based on your isolation requirements and commitment level.

PoolModelBest for
Public PoolShared, multi-tenant, ephemeralExperiments, prototyping, one-off training runs
Managed ClusterDedicated reserved nodes, persistent stateProduction pipelines, continuous retraining, inference serving
Private ClusterYour hardware, orchestrated by ResonTechRegulated data, data residency requirements, existing GPU fleets

See Infrastructure for a full breakdown of each pool type — architecture, data flow, and use cases.

Framework Compatibility

ResonTech runs any standard ML workload without modification. No SDK wrapping. No lock-in.

  • Training frameworks: PyTorch, TensorFlow, JAX, HuggingFace Trainer, Fastai
  • Distributed strategies: DDP, FSDP, DeepSpeed ZeRO, model parallelism
  • Model hubs: HuggingFace Hub, local checkpoints, workspace paths
  • Observability: Weights & Biases, TensorBoard, MLflow — connect your existing tooling
  • Inference APIs: REST, OpenAI-compatible endpoints for LLM serving
  • Runtime: CUDA 11.8+, NVIDIA GPUs from GTX 1080 Ti to H100 SXM5
i
Bring your existing train.py. The kernel handles distributed setup — NCCL configuration, rank assignment, gradient synchronization. Your script sees a standard training environment.