Software and independent advice for private AI

TensorFoundry is a founder-led company in Melbourne. We build software for running AI on hardware you control, and help organisations choose, plan, validate and tune the AI infrastructure they own.

Our Mission

We believe organisations should be able to run AI on infrastructure they control, with the privacy, performance and cost benefits that come with it.

Our software covers the layers of a self-hosted LLM stack: Olla, our open-source LLM proxy, the Alloy gateway, the Pivotal knowledge platform, the Kaizen coding agent and the FoundryOS control plane. We run our own inference platform on FoundryOS in our rack at NextDC M1 Melbourne.

Our consulting practice starts with your workload, not a vendor or a product. We help you choose models and an approach, decide what to run and what to buy, prove what your GPUs actually deliver, and tune the inference you already run, with NVIDIA, AMD and Apple Silicon assessed on their merits. See our consulting services.

Our Values

Privacy First

We believe data privacy is a fundamental right. Our software keeps your data on your hardware, under your control.

Open Source

We're committed to open source. Olla and key components of our stack are freely available for the community.

Performance

We measure before we claim. Published numbers come with the conditions they were measured under, so you can check them on your own hardware.

Developer Experience

Great tools should be a joy to use. We prioritise clear documentation, familiar APIs and helpful error messages.

Meet the Founder

Thushan Fernando, Founder & Principal Engineer

Thushan Fernando

Founder & Principal Engineer

Melbourne, Australia

Thushan is the founder of TensorFoundry and the lead engineer behind Olla, the open-source LLM proxy and load balancer. With 25+ years building software, he specialises in low-level, performance-critical systems work across search, machine learning, and GPU-heavy compute, mostly in Go, Rust and Zig. Previously at SEEK and Readify, where he worked on low-latency search infrastructure and built NeoSearch, an efficient edge-model search platform serving ML at scale. TensorFoundry pivoted to AI infrastructure when the first Llama models landed; the bet was that on-premise inference would matter as much as the models themselves. He also leads TensorFoundry's consulting practice, helping organisations choose, plan, benchmark and tune the AI infrastructure they own. Based in Melbourne, Australia.

  • Low-latency systems
  • Search & information retrieval
  • Machine learning infrastructure
  • GPU compute
  • Inference performance optimisation
  • GPU workload benchmarking
  • NVIDIA, AMD and Apple Silicon
  • Go
  • Rust
  • Zig

Our Journey

TensorFoundry has traded under the name since 2018 and was incorporated as TENSORFOUNDRY PTY LTD in 2026. Its roots are in Sherpa, the founder's ML model management platform, and it moved into LLM infrastructure with Scout (internal platform), an internal on-premises LLM platform, as the first open models arrived.

Those lessons became Olla, our open-source LLM proxy, now past 86.4K+ GHCR container pulls, and now FoundryOS. Every step came from a real infrastructure problem rather than a roadmap slide.

Explore Our Full Journey
2018
2023
2025
2026

Where We Are

Melbourne, Australia

Registered office

Level 1, 470 St Kilda Rd, Melbourne VIC 3004

Serving clients across Australia, New Zealand and remotely.

hello@tensorfoundry.io

TENSORFOUNDRY PTY LTD