Questions, answered straight

We build and optimize production AI infrastructure for engineering teams: cutting LLM costs, adding evals and observability, building MCP servers and internal AI-coding workflows, and making codebases ready for agentic development. Every engagement ships working software, not a slide deck.

Two ways. A fixed engagement of $15–40k, or outcome-based pricing where we keep a percentage of your documented first-year savings. Outcome pricing means we only win when you measurably save — and it usually bypasses procurement. The exact percentage depends on the engagement; we tell you straight on the call.

The cost optimization engagement is 4–6 weeks and targets a documented 30–60% reduction. A codebase readiness audit is 2–3 weeks. Larger platform builds run 6–10 weeks; migrations run 3–6 months.

You should hire — eventually. We help you ship now and hire later, then transition out. We move faster because this is the only thing we build, and our open-source work (brat, harmony-protocol, fast-litellm) is the same infrastructure we deploy for clients.

Yes. We are model- and vendor-agnostic and integrate with what you already run — OpenAI, Anthropic, open-weight models, your CI/CD, Datadog, Linear, and your internal services. We default to the most capable models but route for cost where it makes sense.

A 30-minute technical call to scope the problem, then a fixed-scope proposal. For cost work we start with a read-only audit of your traffic and prompts; for platform work we start with a short discovery against your repo. No long procurement cycle to get started.

Yes. We work under NDA, prefer read-only access for audits, and for regulated teams we can stand up self-hosted inference so nothing leaves your environment. Audit logs and access control are built into every MCP and tooling integration we ship.

Because you pay per token on every single call, and output tokens cost 3–5x more than input (reasoning tokens bill as output too). Cost scales with usage rather than sitting flat like normal infrastructure, so it grows quietly as you add features and traffic. The bill is also usually invisible per route, which means nobody can point at what is driving it — that lack of visibility, not the unit price, is what makes AI spend feel uncontrollable.

Instrument first: tokens, latency, and cost tagged by route and model. Then apply the levers in order — right-size the model per route, cache stable prompt prefixes, batch anything not user-facing, compress bloated prompts, and add fallback chains so timeouts do not trigger expensive retries. Every change sits behind an eval, so a downgrade ships only once you can prove quality held. That combination is how 30–60% comes out with no measurable quality loss.

An API is for your code; MCP is for your agents. Without it, every team hand-rolls its own integration, with its own credentials and no shared audit trail, and each one only works in the tool it was built for. An MCP server exposes a system once as typed tools, and every MCP-capable client — Claude Code, Cursor, and others — can use it under the same auth, scoping, and logging. It turns one-off integrations into governed, reusable infrastructure.

They work at different layers and are complementary. Function calling is the model capability: the model emits a structured request to call a tool, and your code runs it. MCP is the protocol for exposing those tools in the first place, so one server can serve many agents and clients. The model still uses function calling to invoke an MCP tool — MCP just standardizes how the tool got there and who is allowed to call it.

Use a community or official server to prototype — getting a connection working is the easy part. Build (or have us build) when the system is internal, when you need least-privilege credentials, read-only defaults, rate and cost guards, audit logs, and human approval on destructive actions, or when the off-the-shelf server exposes far more surface than you want agents touching. The protocol is a weekend; the governance is the engagement.

Four, and production systems use several at once. Exact or structural match for anything with a right answer (classification, extraction, JSON shape). Rubric scoring for open-ended output against a checklist. LLM-as-judge for nuance, with the judge itself validated against human labels. And human review, reserved for the highest-stakes samples. Offline evals run in CI to catch regressions; online monitoring scores real traffic and feeds new cases back into the suite.

An agent gets a task most of the way there fast, then the last stretch — the edge cases, the integration, the part that has to actually be right — takes longer to review and repair than writing it would have. It is real, but it is mostly a property of the environment rather than the model: agents thrash when there is no feedback signal. Fast tests, strict types, clear module boundaries, and a good context file shrink that last 20% dramatically.

Score it on five axes: module boundaries, test coverage and speed, type strictness, written context (a CLAUDE.md or AGENTS.md file), and whether agents can reach internal systems over MCP. The cheapest high-return fixes are almost always the same three: write the context file, make the test suite fast and runnable, and turn on strict types. The expensive structural work comes later, and only where the score says it will pay off.

Yes. We are based in Scotland and work with engineering teams across the UK, the EU, and the US. Engagements are remote by default with overlap hours agreed up front — UK and EU teams get full-day overlap, US East Coast gets a solid afternoon, US West Coast gets a morning block. For regulated EU and UK work we can keep data in-region, including standing up self-hosted inference so nothing leaves your environment.

A read-only audit — typically two to three weeks — is the usual front door. It gets both of us to numbers before anyone commits to a build: what your AI actually costs by route, or how ready your codebase really is. Most larger builds start there, and you are free to take the report and do the work yourself.

Then you probably need less than you think, and we will tell you so. The common shape is that we do the specialist piece your team has not had time to own — the cost audit, the eval harness, the first MCP servers — set the patterns, and hand it over. We would rather do six good weeks and transition out than become a dependency.

Yes. We work under NDA as standard, prefer read-only access for audits, and follow your access, review, and data-handling rules. For sensitive environments we can work entirely inside your infrastructure, and for regulated data we stand up self-hosted inference so nothing sensitive leaves your perimeter.