Taking on new engagements
Hard systems made shippable.
KaijuTank is a founder-led engineering studio. We build AI systems that run on your own hardware or your cloud, delivery pipelines with real gates, and the platforms around them, then take the work from prototype to production.
Measured in our own work
Figures are from KaijuTank’s internal test suites and benchmarks, October 2026. Inference speed is burst performance on an RTX 3080 Laptop GPU.
Start here
Know where your cloud delivery will break, before it does.
The CloudLab Delivery Assessment is a fixed-scope review of how your system gets built, tested, released and handed over. It suits a team that is inheriting a platform, preparing to hand one to a client, or shipping more often than its pipeline was built for.
You get written findings you can act on with or without us. If you then want the fixes built, the assessment is the scope for that work.
CloudLab Delivery Assessment
Fixed fee rate sheet on request, quoted in writing first
- Cloud readiness. How environments are built and promoted, and what is not reproducible.
- CI/CD risk. Where a bad change can reach production, and which gates are missing.
- Handoff quality. Whether another team could run it from what is written down.
- A modernisation roadmap. Ranked fixes, with the first ones small enough to start on at once.
Services
Senior engineering where AI, cloud and product meet.
Bring us in when a project crosses the boundaries between product software, AI systems, infrastructure and release automation, and needs one accountable engineer rather than four vendors.
AI and agent systems
Model selection and routing, retrieval, structured tool use, memory, evaluation and agent workflows, designed to run privately on local hardware or in your cloud.
- Private / local AI
- Agent architecture
- Evaluation and guardrails
Cloud delivery and CI/CD
Infrastructure as code and release pipelines built for repeatability: environment promotion, security and policy gates, approvals, observability and a record of what shipped.
- Terraform / Docker
- GitHub Actions / Jenkins
- Approval-gated release
Product and platform engineering
Architecture and implementation for browser products, desktop apps, internal platforms, developer tools and backend services, including the unglamorous work of making a prototype reliable.
- Product architecture
- Desktop and web apps
- Systems integration
Creative and real-time automation
Technical workflows for Unreal Engine, generative media, 3D assets, rendering and review, plus automated QA that drives the real application instead of a mock.
- Unreal workflows
- AI media pipelines
- Runtime QA
Plans
Three ways to work with us.
Start small and scope it in writing. Every engagement ends with something you can run, read or hand to your own team.
Assessment
Start hereCloudLab Delivery Assessment
Fixed fee Rates on request, quoted in writing
You want to know where your cloud delivery is at risk before you change it.
- Review of how the system is built, tested, released and handed over
- Written findings with risks ranked
- Cloud readiness and CI/CD risk report
- A modernisation roadmap you can act on with or without us
- Scope and fee agreed in writing before work starts
Build
AI / Cloud Implementation Sprint
Per sprint Rates on request, quoted in writing
You have a defined thing to build and want it delivered, tested and documented.
- One defined deliverable: a pipeline, service, model deployment or product milestone
- Acceptance checks agreed up front
- Delivered with tests and documentation
- Built with our own tooling: CloudLab, Aether, Runtime QA Agent
- Runbooks so your team can own it afterwards
Ongoing partner
Fractional Principal Engineering
Monthly Rates on request, quoted in writing
You need senior hands on architecture and delivery without hiring a full-time lead.
- Architecture decisions and design review
- Hard debugging alongside your engineers
- Delivery support from prototype to production
- One point of contact, no account managers
- Month to month
We send our rate sheet on request: ask for it by email. Every engagement is quoted in writing before any work begins. Product access (previews, early access, design partnerships) is by request, and product pricing will be published when each one is generally available. See what you can ask for on each product.
How we work
Every change earns its way to production.
We run our own products through the same gated pipeline we build for clients. Automation does the checking and the first attempt at a fix; a person decides what ships.
- 1
Check
Each project declares its own tests. They run on every change and again every night.
- 2
Locate
A failure is traced to the code responsible before anyone reads a whole file.
- 3
Patch
The smallest change that addresses the cause, made on its own branch.
- 4
Verify
The project’s checks must pass again. A fix that fails is rolled back, not merged.
- 5
Release
Nothing is published, deployed or pushed without a person’s approval.
Evidence over assertion
Claims come with the test run, the benchmark or the log that supports them.
Local first
If it can be validated on one machine before it touches a cloud account, it is.
Small, reversible steps
One change, one branch, one verification. Easy to review and easy to undo.
Yours to keep
Work is delivered with the runbooks and context your team needs to own it.
Featured platform
CloudLab turns cloud delivery into a controlled, evidence-aware system.
CloudLab is KaijuTank’s local-first cloud delivery and client handoff platform. Work progresses from local → development → sandbox → production through promotion gates and approvals, with infrastructure as code, observability and a traceable record of every step.
Fourteen workspaces cover it end to end: the pipeline and its environments, releases with promotion and rollback history, an architecture canvas and component inventory, security controls and policy checks, monitoring, logs and alerts, and the governance side: a traceability ledger, an audit timeline and access roles. A step can be run by hand or with an agent.
The browser experience is built. The Windows and Linux desktop app is in final look-and-usability testing.
KaijuTank Labs
The products behind the practice.
Client work and product work feed each other. These are the systems we build and run ourselves; the architecture, tooling and lessons carry straight into engagements.
CloudLab
Cloud delivery + handoff platform
Local-first validation, environment promotion, approval gates and an evidence-aware handoff package for production cloud systems.
Aether
Fine-tuned language model
A language model we fine-tune, evaluate and release ourselves. It plans, calls tools and returns structured output inside the products that use it, served entirely on local hardware.
Silo
Inference engine + compression research
A Rust engine for hardware-aware model serving on small GPUs, with verified streaming compression, KV-cache planning and CUDA work underneath.
Cinder
AI-native browser + Creative Studio
A desktop workspace that brings research, building, automation and image, video, audio and 3D creation into one application with local models.
Frameward
Production tracking + creative compute
A production platform for film, VFX, animation and games: tracking, immutable publishing, asset lineage and review, with Unreal, Unity, Blender and Godot integrations.
Borealis
Unreal Engine AI agent for VS Code
Generates Blueprints from plain language, automates asset pipelines and routes work across local and cloud models, with a headless mode for CI.
KaijuCode
Agentic developer runtime
Local-first tool execution with model routing, sandboxed file and shell access, memory and context compression for long-running engineering tasks.
Runtime QA Agent
Embodied software QA
A Rust runtime that operates real desktop applications and games, records every observation and action for deterministic replay, and files evidence-backed findings.
Interview Coach
Technical interview practice
A desktop coach for practice and mock interviews that follows the problem, the code and the conversation, with local speech recognition and local models.
EdgeLab
Evidence-first strategy research
Turns a trading hypothesis into backtests, walk-forward validation, sensitivity analysis and an audit report. Research and paper trading only; it makes no profit claims.
Kaiju Engine
Engine and editor research
A Rust engine core with a web-technology editor, scripts in several languages compiled to sandboxed modules, and a stable C interface.
Inside the work
Built as systems, not feature lists.
Two examples of how the pieces fit together.
- Base model
- Fine-tune on our own data
- Evaluation and quality gates
- Release a version
- Served locally by Silo
A fine-tuned model, served on an 8 GB laptop GPU.
Aether is trained, evaluated and released through explicit quality gates, then served by Silo. We benchmark it the way an agent actually uses it: how fast it reads a long prompt, not only how fast it writes.
A desktop product with thousands of automated checks.
Cinder combines a browser, a build workspace and a creative studio. Its test suite runs on every change, and a nightly run across our portfolio attempts a verified fix when something breaks.
Founder-led
Principal-level engineering without layers of handoff.
KaijuTank is led by Mark Salmon, a software and AI systems engineer with experience across real-time production, cloud infrastructure, developer tooling, CI/CD, Android and embedded systems, and client-facing technical leadership.
The person you talk to is the person who does the work. The company operates as a focused engineering partner while building reusable platforms through KaijuTank Labs.
Start a conversation
Have a hard systems problem?
Bring the cloud delivery problem, the AI idea, the failing pipeline or the prototype that needs to become a product. A short note is enough to start: what you are building, what is in the way, and when you need it.
