← Bobby Mathews

Synkti

70% Cost Reduction for AI Inference

Run AI Models for a Fraction of the Cost

Synkti orchestrates spot instances—volatile cloud servers that cost 70-80% less than on-demand. Our preemption prediction and optimal migration means your inference stays reliable while costs drop.

How It Works

1. Predict Preemptions

Spot instances can be interrupted with 2 minutes warning. Synkti forecasts these events using historical patterns, migrating workloads before interruption occurs.

2. Optimal Migration

When migration is needed, we use the Kuhn-Munkres algorithm to assign workloads to standby instances—minimizing disruption and maximizing resource utilization.

3. Stateless Failover

During the 2-minute warning window, we gracefully drain active requests and spawn replacement instances. Clients retry naturally via HTTP, and new instances load models from cache. This is 6x faster than checkpoint migration with zero GPU memory serialization.

Read why we chose stateless over checkpoint migration →

4. P2P Architecture

No central control plane to fail, scale, or pay for. Each node runs its own orchestrator and discovers peers dynamically. Like a flock of birds—no leader, yet perfectly coordinated. Self-aware, self-monitoring, self-healing.

AWS uses EC2 tag-based discovery (Phase 2). DePIN uses libp2p (Phase 3).

Validated Results

70-80% Cost Reduction
2,191 Lines of Rust
32 Tests Passing

Try the Simulation

Don't take our word for it. The simulation engine validates cost reduction strategies against real spot market data. Clone and run it yourself—no commitment required.

Technical Depth

Built by Bobby Mathews, a Rust systems engineer specializing in distributed infrastructure. Synkti uses a discrete-event simulator with 243-scenario validation, implementing research from SpotServe (OSDI 2024) and SkyServe (EuroSys 2025).

Learn more about the engineer behind Synkti →

Roadmap

Phase 1: Research Prototype
  • Kuhn-Munkres optimal migration algorithm
  • Grace-period exploitation (drain & respawn)
  • Discrete-event simulation engine
  • 243-scenario validation matrix
Phase 2: P2P Production Orchestrator
  • P2P architecture (each node self-governing)
  • EC2 tag-based peer discovery
  • Stateless failover (Drain → Select → Spawn → Route)
  • AWS integration (SSM, ELB, Spot API)
  • Pilot validation with early adopters
Phase 3: Global P2P Network
  • libp2p peer discovery (mDNS + Kademlia DHT)
  • Multi-cloud support (AWS, GCP, Azure, bare metal)
  • Solana on-chain settlement layer
  • Permissionless node participation

Read the full vision and decentralization thesis →