LLM apps with retrieval
RAG systems that handle long context, multi-tenant data, and citation-quality grounding. Eval-driven, not vibes-driven.
~ $cat./lab.md
MRA AI LabAI Engineering.
Shipped, not deck-ware.
A small senior team that builds production AI on Google Cloud for regulated industries. Priced per build, not per hour.
~ $cat./engagements
Three engagement modes.
Match the unit of risk to the unit of trust. Start small, expand only when the previous step proved out.
discover.ts▶ ready
Discover
2 weeks · fixed
Feasibility study, eval harness, and a working demo on a small slice of your data. Decide whether to invest further with evidence in hand.
- Vertex AI
- Gemini
- Vertex Eval
build.ts▶ running
Build
6–12 weeks · scoped
A production system in your GCP project: terraform, CI/CD, evals, observability, and a runbook your team can operate.
- Cloud Run
- BigQuery
- Workload Identity
operate.ts▶ deployed
Operate
Quarterly · capacity
Reserved engineering capacity for ongoing improvements, model upgrades, and production incidents. We stay until you do not need us.
- On-call
- Eval pipelines
- Model rollouts
~ $cat./what-we-ship
What we ship.
A focused list. We say no to most things so we can do these well.
Agentic workflows
Tool-using agents with budgeted reasoning, human checkpoints, and structured handoff. Built on Agent Builder + custom orchestration.
Domain-tuned models
Supervised fine-tunes on Gemini and open-weights, with held-out evals you can interrogate and version.
Eval & observability
We treat evals as first-class infra. Every model change ships with a side-by-side scorecard, regression tests, and prompt diffs.
Production hardening
Rate limits, retries, fallbacks, secret rotation, audit logging. Real systems break — yours must not break first.
Compliance posture
Workload Identity, CMEK, VPC-SC, DLP, and Canadian data residency. We start with security, not bolt it on.
~ $cat./past-engagements
A sample of past work.
Anonymised by design — we expect our clients to ask for the same. Reach out for references.
RegTech for a Tier-2 fintech
Built an end-to-end KYC + sanctions co-pilot in 8 weeks. Cut analyst time per case from 14 minutes to 2:30.
Trade-doc parsing for a Pacific importer
Document AI + Gemini extracts customs forms across 6 languages. 96% field-level accuracy, with human review on edge cases.
AML co-pilot for a digital-asset operator
Agentic case builder shipped to production with full audit trail. Passed two regulator examinations on first ask.
Got a problem worth a Discover?
Two weeks, one fixed price, one written deliverable. The worst case is you walk away with a working demo and a clear-eyed recommendation.