Union Square Consulting, homeBook a 30-minute callMenu

Embedded AI engineering · London · Dubai · Delaware

AI in production, built inside your teams.

Our principals lead every engagement, with senior engineers working alongside your people to build AI agents that know your business, checks that stop their mistakes going live, and pipelines that recover on their own. The systems we built for a 1,000-person agency have run in production since April 2026.

Your data is organised into business knowledge, AI agents work from that knowledge, and in what we build, automated checks test their output before it reaches live systems.

Fig. 1 How the work flows
Fig. 1 How the work flows

Your data is organised into business knowledge, AI agents work from that knowledge, and in what we build, automated checks test their output before it reaches live systems.

Our principals have delivered inside Meta, Spotify, UBS, Starling Bank, S&P Global, Sky Italia, Viasat, Brainlabs, Société Générale and the European Investment Bank.

01 Approach

Most enterprise AI dies between the demo and the day shift. We build the other kind from inside your teams, designed for production from day one and handed over so your people run it without us.

The usual options, and ours

How a Big-4 programme, staff augmentation and USQ compare on who builds, how the work starts, the timeline and what you keep.
Who buildsHow it startsTimelineWhat you keep
Big-4 programmeJunior consultantsA partner’s pitch6–18 months to production, if it gets thereA roadmap
Staff augmentationIndividual contractorsCVs for you to sift throughDepends on your management bandwidthWhatever your systems could already support
USQForward deployment unitHands-on principals, plus senior engineers as neededEmbedded with your teams from week oneProof in weeks 2–4, production-ready in months 2–3A live system, plus playbooks and in-house champions

02 In numbers

Selected figures

Adoption in North America
91%
Brainlabs, Anthropic customer story
At-risk renewals and upgrades the tool covered
$50M
Inmarsat sales tool, built on usage data
Conversations scored to compare AI releases
100,000+
Meta AI for Customer Support, quality scoring
Subsidiary ERPs moving onto one NetSuite
40
Brainlabs, in progress, with data errors caught in testing before go-live

03 Credentials

Two credentials, and no favourite model.

We build on Claude, GPT and Gemini and pick the right model for each job, so neither credential decides which one you use.

  • Partner network

    OpenAI Partner

    Union Square Consulting is a partner in the OpenAI Partner Network.

  • Customer story

    Published by Anthropic

    Anthropic published the Brainlabs deployment as a customer story, including the 400 or so skills (reusable instructions for Claude) that employees wrote in the first four weeks. Our principals built the platform those skills run on, along with the shared business memory and data pipelines it depends on.

    claude.com/customers/brainlabs (opens in a new tab)

“If you’re doing something more than three times, you should try to automate it.”

Ben Vincent, CTO, BrainlabsFrom Anthropic’s customer story

04 Capabilities

Production AI is three engineering problems.

The AI agents, the business knowledge they draw on and the systems they run on fail in different ways, so each has a principal accountable for it, and the team around it grows with the engagement.

  1. AI adoption & agent architecture

    Fig. 1 03 AI agents 04 Checks

    Led byPrincipal Architect

    We design how AI agents work across a whole organisation, from access rights and platform rules down to the automated checks the agents’ work has to pass. Adoption is part of the build, because a platform nobody uses is just a cost line.

    Evidence

    • Production OAuth 2.0 MCP server in Go, built solo in five days.
    • Verifier-gated agents compressed multi-week migration work into controlled 2–3 day cycles.
    • Claude access provisioned by job function, with drift correction.
  2. Data platform & business knowledge

    Fig. 1 01 Your data 02 Knowledge

    Led byPrincipal Engineer

    An AI agent that starts every task from a blank page has to rediscover what your business already knows. We build the shared memory that hands it the relevant conversations and account context, and the data platform underneath.

    Evidence

    • 12× throughput on a rebuilt data-integration platform, ETL runtimes down about 90%.
    • Snowflake compute and storage fell 87%, and about $36k a year of SaaS licensing came in-house.
  3. Infrastructure & reliability

    Fig. 1 05 Live systems

    Led byPrincipal Data Engineer

    This layer decides whether an AI system is still running next quarter. We build the scheduling, testing and release automation underneath your agents and keep at it until the pipeline can recover from failures without anyone firefighting.

    Evidence

    • Starling Bank’s low-latency fraud-inference integration was built to a five-nines availability target.
    • AWS data lake with real-time and batch ingestion for S&P Global Platts.
    • Azure migration assessment for UBS, with Kafka Streams, Cosmos DB and Kubernetes.

No workstream hangs off one person.

Every workstream is covered by at least two principals.

Capability coverage by principal. Core means a core strength. Working means solid working capability.
Principal ArchitectPrincipal EngineerPrincipal Data Engineer
Custom agent harnessesCoreCoreCore
LLM & agent systemsCoreCoreCore
RAG, memory & context engineeringCoreCoreCore
AI evaluation & verificationCoreCoreCore
Prompt engineering & skills authoringCoreWorkingCore
Claude Code & platform governanceCoreWorkingWorking
Data platforms & warehousingWorkingCoreCore
CI/CD, cloud & infrastructureWorkingCoreCore
React / front-end deliveryCoreCoreCore
C-suite & stakeholder managementCoreCoreCore

05 How we deploy

From first workflow to steady state.

The phases draw on years of production data work in banks and consumer platforms, and they move quickly because the expensive mistakes already happened somewhere else.

  1. Weeks 1–2

    Embed and map

    We sit inside your teams and map the workflows, constraints, data and AI maturity you actually have.

    Fig. 1 01 Your data

    You leave withUse cases ranked by value, ready to show a CFO

  2. Weeks 2–4

    Prove fast

    A first working version runs in conditions close to live use, judged against criteria agreed before the build starts.

    You leave withEvidence for a build or no-build decision

  3. Months 2–3

    Make it production-ready

    Automated checks on the AI’s work go in alongside controls on who can access what, long before scale forces the issue.

    Fig. 1 04 Checks

    You leave withA governed system before anyone scales it

  4. Ongoing

    Hand over

    Workshops, playbooks, reusable AI skills and local champions put the work in your people’s hands, so you stop renting it from us.

    You leave withYour own people building the automations

  5. Steady state

    Optimise

    We audit what your AI actually costs, then tune cost and speed ahead of demand, so neither the bill nor the response times catch anyone out.

    You leave withCost and speed headroom before you need it

07 Principals

You meet the people who build.

We came up through data, platform and infrastructure engineering before AI became the job, so the pipelines get as much attention as the model. The principals lead every engagement and stay on the build, bringing in senior engineers from a wider bench as the work needs.

Principal Architect

10+ years

AI adoption, agent systems and Claude architecture

The Principal Architect works from C-suite advisory down to the hard-coded controls agents cannot talk their way past. The platform this principal founded now spans about 800 skills across 82 plugins, and its founder is still the lead architect.

Selected record

  • Cut Claude slide-deck prep from 50 minutes to 5; routing and context engineering lifted efficiency 74% estate-wide.
  • Primary author of 18 Claude Code subagents and 20 of 24 production skills on a financial data warehouse.
  • In independent consulting, delivered systems associated with a 46% cut in first-response time and 82% less manual triage.

Stack

Python · Go · TypeScript · MCP · Multi-agent systems · RAG · Snowflake · BigQuery · Cloud Run · Claude/GPT/Gemini

Principal Engineer

14 years

Data platform, context and memory architecture

A former Staff Engineer who has also led a data team of ten engineers and analysts, the Principal Engineer now designs the context that agents start work from.

Selected record

  • Designed the data models and built the pipelines for Meta’s GenAI evaluation, under FTC/EU-grade PII controls.
  • Moved Inmarsat’s dbt scheduling and PR checks to GitHub Actions when dbt Cloud’s per-seat pricing did not add up.
  • Led Edvantage’s platform as fractional CTO, with AI grading that sends low-confidence submissions to review.

Stack

Python (asyncio, FastAPI) · dbt · Airflow · Snowflake · BigQuery · Databricks/Spark · Beam/Scio · GCP/AWS/Azure · Terraform · React/Next.js

Principal Data Engineer

17 years

Pipeline orchestration, CI/CD and data infrastructure

Since 2009 the Principal Data Engineer has worked client-side in banking, fintech, media and energy, on the data pipelines and release automation that everything downstream depends on.

Selected record

  • Architected Airflow orchestration, then corrected the CI/CD until a self-healing pipeline could run under Claude managed agents.
  • Spent four years embedded as a senior data engineer at Sky Italia, one of Europe’s largest broadcasters.
  • Delivered a data analysis and economics reporting platform for the European Investment Bank.

Stack

Airflow · Python · Java · SQL/Postgres · Kafka/Flink/Spark · AWS/Azure/GCP · Snowflake-pattern warehousing · CI/CD · Docker

The Principal Engineer and Principal Data Engineer have worked as a unit for nine years. References on request.

08 Terms

Procurement should be the easy part.

Contract
One supplier and one contract, on fixed-scope statements of work.
Staffing
Senior engineers only, led by a principal who stays on the work, with no bait-and-switch.
Control
We build inside your systems, under your access controls.
Contact
One accountable principal as your point of contact, however large the team gets.
Continuity
If one workstream surges, we add senior engineers to it, so nothing stalls on a single hire.
Scaling
The team and the commercial terms flex as the scope changes.

09 Next step

Bring one workflow. Get a straight answer.

In 30 minutes we will tell you how we would take it to production and what we would leave alone. If the honest answer is “don’t build it”, that is the answer you get.

If the calendar does not load, open the booking page (opens in a new tab).