AI Platforms and Models We Build On
Software Depo is a specialist production-AI engineering firm. We design and ship AI agents, MCP servers, and retrieval systems on the platforms and models that matter for real business work — chosen on the task in front of us, not on hype.
Our engineers work across frontier models, agent and orchestration frameworks, the Model Context Protocol, retrieval and vector infrastructure, and enterprise AI clouds — always with guardrails, evaluation, and human oversight built in.
Platforms and Models
AI Platform Engineering
Software Depo builds on frontier models (Claude, GPT), the Model Context Protocol, Copilot Studio, retrieval and vector infrastructure, and enterprise AI clouds — engineered with scoped permissions, evaluation, and human approval. Explore our AI development services or tell us what you want your AI to do.
How a Model Gets Chosen for a Job
Ranking models in the abstract is a poor use of time. What matters is which constraints bind on the job in front of you. Long-context reasoning over a large document set, disciplined tool calling, latency under load, and code generation quality pull in different directions, and a model that leads on one can lag badly on another.
Then come the constraints that have nothing to do with capability: whether the model is available in a cloud region your policy allows, whether your legal team can live with the data-handling terms, and whether you can pin a version and know how long that version stays available.
Keeping the Model Swappable
Model choice is more reversible than most of the architecture around it, but only if the code is written for it. Keep prompts, tool schemas, retrieval logic, and evaluation sets out of any single vendor SDK and a swap becomes something you can test. It will not be free. Behavior differs between models even when the interface does not, so expect prompt re-tuning and a full re-run of your evaluation set. What the abstraction buys you is a swap you can measure instead of a rebuild you have to schedule. Skip the discipline and the model quietly becomes your architecture.
The Plumbing You Live With Longer Than the Model
Where identity comes from, which credential each tool call runs under, where the audit log lands, what happens when a tool call fails halfway through a multi-step task: all of this outlasts whichever model is current. It also determines whether anyone outside engineering can answer for what an agent did last Tuesday.
Platform shape follows from that. An organization already deep in Microsoft 365, automating work close to Graph data, has a reasonable case for looking at Microsoft Copilot Studio first, because its governance surfaces sit in admin tooling the security team may already use. Work needing unusual control over retrieval, orchestration, or self-hosting fits a framework build better. The platform you already pay for is sometimes the right answer.
Model Updates Are Dependency Updates
Frontier models change underneath you. Behavior shifts, sometimes in ways a person notices before any metric does. Without an evaluation set you hear about it from a user. Pin versions where the platform allows it, keep a regression set that reflects your actual traffic, and put a model change through it with a rollback path ready.
Private, Hybrid, and On-Premise
When data cannot leave your environment, options run from a dedicated cloud tenant with contractual boundaries through to open-weights models on hardware you own. Self-hosting moves real work onto your team: capacity planning, patching, model updates, GPU availability, and a narrower field of models to choose among.
Before anyone commits to hardware, find the written policy or the regulatory obligation that forces it and read the actual text. If nobody can point at the rule by name, a hybrid arrangement with clear data boundaries deserves a look first.
Vendor Names on This Site
The platform pages here are written from the outside. Anthropic Claude, OpenAI GPT and Codex, Microsoft Copilot Studio, Model Context Protocol, Retrieval and Vector Databases, AI Skills and Capabilities, and AI for Niagara and Operations each name the technology that page discusses, and nothing beyond that. Naming a product is not a partnership, a certification, or an authorization, and none is claimed here. Product names and trademarks belong to their owners.
Vendor terms, region availability, and model lineups move faster than a page like this can. Where a specific claim matters to a decision, confirm it against the vendor’s current documentation before you act on it.