Simas Razinskas

LLM systems

The machinery around the model call: gateways across providers, GPU inference orchestration, prompt optimization, and the fallbacks for the day an API goes down.

For teams past the prompt stage, now facing service boundaries, fallback paths, observability, cost control and evaluation.

Case studies behind this

  • A multi-provider AI gateway with built-in cost accounting

    Direct model-provider calls spread SDK quirks, rate limits, failures, and pricing logic across feature code.

    Go adapters · gRPC / protobuf · persisted in-memory queue
  • Orchestrating long-running inference on serverless GPUs

    Long-running image-generation jobs could lose work or leak GPU memory when a WebSocket dropped or an ephemeral worker disappeared.

    RunPod · ComfyUI · WebSockets · S3 offload
  • Evolutionary prompt optimization

    Prompt tweaks changed output quality unpredictably, and manual testing could not explore enough variants to find better prompts reliably.

    BLEURT · LLM APIs · evolutionary operators · NumPy
  • A local-first dictation pipeline for macOS

    Cloud dictation was not acceptable for sensitive work, while context-free local transcription produced text disconnected from the current screen.

    Swift / SwiftUI · whisper.cpp · Accessibility & CoreGraphics APIs

Questions worth asking first

  • How will latency, cost and fallback behavior be visible in production?

  • Which provider or model boundary should the product depend on?

  • Do we need RAG, fine-tuning, workflow design — or just a simpler prompt boundary?

Related services

LLM productionization

Start here when the prototype is convincing but nobody can say what it costs at scale, why it fails, or whether yesterday's prompt change made it worse.

Open service

Have a system that has to work?

Book a call and we'll map it out: what it should do, where it will break, and the smallest version worth building first.