We diagnose before
we prescribe.
No system we've built has reached production without a pass/fail eval score. That isn't a policy, it's the only responsible way to deploy AI into a business that depends on it.
Most AI systems are
built backwards.
The model is picked first, not last. The eval is written after the fact, if at all. The deployment is optimistic. We invert all three.
No eval framework
The system ships with a demo that worked. Then it meets real data, edge cases and compliance, and fails 15% of prompts with no one willing to sign off. There was never a definition of "working."
No production architecture
A notebook that calls GPT is not a production system. Without MCP boundaries, observability and graceful fallbacks, the first unexpected input breaks everything.
No regression protection
The model improves, the prompt changes, the data shifts. Six months on, accuracy has quietly dropped 20% and no one noticed, there was no CI gate to catch it.
No surprises.
No black boxes.
Every phase has a deliverable you review and approve before the next begins. The eval harness is built in parallel with the system, never bolted on after.
Define the automation boundary, agree the success criterion, confirm timeline.
Focused build in your stack, QA included. You review a working demo, not a slide deck.
Production deployment, self-hosted, full ownership transferred on day one.
Match the model
to the problem.
Both include full ownership, eval testing and production deployment. They differ in scope, timeline, and what "production" means for the problem.
Engineering Program
Custom architecture, full engineering practice.
- Full discovery & requirements phase
- System architecture document, approved before build
- Multi-agent system or agentic pipeline from scratch
- Complete eval harness, must pass before production
- CI regression gate in your pipeline
- Full observability: LiteLLM + Langfuse from day one
- Architecture docs + operational runbook
- 30-day post-handoff support
Agentix Rapid
One targeted automation, self-hosted, live in days.
- Scope call to define the automation boundary
- Single-focus automation built in 2–5 days
- QA and functional testing before delivery
- Production deployment to your infrastructure
- Self-hosted, your data never leaves your environment
- Full code ownership on delivery
Eight things that
are never optional.
These apply to every system, in every engagement, at every size. They are what separate a prototype from production software.
The standards aren't a manifesto, they're measured. What that rigor produces, in named systems and for named clients:
- £25K/yr receptionist cost replaced for A&A by a Voice AI agent, plus ~20 hrs/week of staff time saved.
- 12h → <5min VistaGPT cut Power BI PBIR report builds from twelve hours to under five minutes across 5 production clients.
- ~3,953 commits across 26+ repos, 4 production systems plus 1 in active development. This is depth, not a deck.
Every merge is blocked if it degrades performance, preventing the slow "works on my machine" drift that kills AI systems.
Orions: CI blocks PR if precision drops >5pp or ECE >0.15
Deterministic retrieval and schema constraints enforce grounding before any LLM writes. Architectural, not prompt-based.
VistaGPT: 6 parallel ChromaDB stages + reconcile before SQL
Every agentic system has defined checkpoints where a human confirms. Bypass is an explicit, named flag, never the default.
TAFI: 3 named gates · AI CompMate: Observer → Partner ladder
Traces, per-stage latency, cost-per-query and error rates, instrumented before launch, not when it breaks at 2am.
LiteLLM + Langfuse across all systems · Prometheus + OTel on Orions v2
New versions run in shadow mode beside v1. You compare outputs on real traffic before cutting over. No surprise risk.
Orions v2: DIAGNOSIS_V2 flag · compare via GET /eval/shadow-report
You own it
from day one.
Every system is delivered with the documentation and infrastructure for your team to operate, extend and maintain it, without us.
Start a project →- Production deployment, verified working, monitored from day one
- Complete eval test suite, runnable by your team at any time
- CI regression gate, configured in your pipeline, blocks quality drops
- Architecture documentation, system diagram, components, decision log
- Operational runbook, how to monitor, debug and extend
- LiteLLM + Langfuse observability, dashboards set up and accessible
- 30-day support window, direct access for questions and edge cases
- Full code ownership, every line committed to your repo from day one
The questions worth asking.
Do I need to provide a technical spec?
No. The Discovery phase is where we build the specification together. You explain the problem, what you're automating, what the failure mode is, what success looks like, and we write the architecture doc from there. If you already have a spec, even better.
What happens if the system doesn't pass eval?
It doesn't ship. The eval harness defines what "working" means for your system; if it doesn't pass, we fix it, that's part of the build, not a separate cost. TAFI's E13 story is the clearest example: the eval showed QA 100, but platform verification found 5 failure classes it hadn't covered. We fixed all 5 before shipping.
Can I host this myself?
Yes, and we prefer it. Self-hosted means you own the data, model calls don't route through any third-party logging layer, and you're not paying us usage fees indefinitely. We deploy to K3s on Hetzner for most systems, but can target any Kubernetes-compatible environment.
Do you offer retainers or ongoing support?
A 30-day support window comes with every engagement. Beyond that we work on a project basis, a new scope call for new work. The best relationship is a client who owns their system and returns to extend it, not one paying a monthly fee for access.
What's the smallest project you take on?
A Rapid Automation, a single, well-defined process, 2–5 days. The constraint isn't budget; it's that the problem must be specific enough to automate in that window. The 30-minute scope call will tell you whether yours fits.
Tell us about
your project.
Describe the process you want to automate, the data you want to query, or the problem you want to solve. We'll tell you how we'd build it, what it would take, and the eval it would have to pass.