Family-built, engineer-led lab distilling sparse intelligence into production-grade systems with transparent data attribution - tailored to tight inference SLOs.
Droidcraft OY is a family-owned research lab of hands-on engineers, systems operators, and scientific compute specialists. We design, deploy, and operate sparse expert architectures that preserve transparent data attribution and privacy assurances with a on-prem / private cloud focus, while sustaining the inference SLOs demanded by live operational systems at scale.
Every artifact we ship - from kernels to orchestration - remains type-safe, free from license burdens, and maintainable long after the pilot wraps.
Designing and training of task-specific SMoE architectures, with transparent data attribution, and the ability to merge experts from separate training runs. We will quantize and tune for low-latency inference as demanded by applications.
End-to-end experiment and pilot design, evaluation and training harnesses, and statistical analysis that increase model transparency before they reach regulated workloads.
Custom kernels, sharding strategies, and throughput tuning validated on H100, H200, MI250X, and federated K8s clusters to push utilization past industry benchmarks.
Formally quality-gated codebases, audit trails, and restriction-free license stacks that satisfy cross-industry governance, privacy, and security reviews.
Sparse-mixture-of-experts transformer crafted to validate our zero token-drop router, expert merging, 8-bit optimizer workflow, and full training pipeline end to end.
Industry-specific post-training of a 70B Llama 3.2 variant with IPO objectives and FP8 quantization for single-superchip deployment.
Latency-tuned inference stack on ARM-based Grace Hopper nodes pairing custom Triton kernels with streaming adapters for hybrid clouds.
Design and integration of application-specific MCP servers wired into standard LLM agent pipelines with transparent routing and audit hooks.
Implementing private model architectures on top of vLLM and SGLang with observability, structured output tooling, and safe tool-calling. Maintainers of an Anthropic feature-complete API for SGLang.