Are you an AI consultancy, or a data engineering consultancy doing AI?

The second. The honest read is that most “AI projects” are 80% data engineering and 20% model work, and most “AI consultancies” are PowerPoint shops that subcontract the data engineering. We've been doing the data engineering underneath production systems for 15 years; AI just changed what the output layer is called.

Will you fine-tune a foundation model for us?

Almost never, and we'll usually talk you out of it. Most use cases that founders think need fine-tuning are actually retrieval problems, prompt design problems, or data quality problems. We'll tell you in the audit whether yours genuinely needs fine-tuning; if so, we'll partner with a model-shop or run it ourselves depending on the shape.

How do you handle RAG and retrieval over messy data?

Entity resolution and normalisation before anything else. A retrieval system over duplicate, inconsistent records gives the model confidently-wrong answers, which is worse than refusing. We layer pg_trgm for the bulk of the matching, Jaro-Winkler for short identifiers, and vector embeddings only where the problem is genuinely semantic.

What about evals and regression testing for agents?

Every engagement ships with an evaluation harness, a golden test set, regression suites, and a pass-rate dashboard. We don't deploy agents that regress against the golden set. “It felt better in vibes” is not a release criterion.

Can you operate in regulated environments (finance, health, compliance)?

Yes. We built a sanctions-compliance platform for a global financial-crime advisory that runs in a heavily-regulated environment across 20+ financial institutions; the Princeton Research Accelerator built IRB-compliant research data governance. AI work in regulated industries is mostly about lineage, audit trails, and access control, which is the unsexy data engineering we lead with anyway.

How long does an AI engagement typically take?

Eight to twelve weeks for data audit through deployed agent serving real traffic, the standard shape. Longer if the audit reveals that the data work has to come first (which it often does, and we’d rather you know that early). Phased and milestone-bound; you can stop at any phase boundary.

Capability · Agentic Workflows & Machine Learning

AI that works in production, not just notebooks.

Most AI projects die in the data layer, not the model layer. We do the unsexy work underneath, entity resolution, governance, pipelines, evals, so the agentic systems and ML you ship hold up against real traffic, not just notebooks.

Take the AI-Readiness Audit Talk to our team

Data quality first · Evals before launches · Production-grade infra · Knowledge transfer baked in

Why AI projects fail

It’s almost never the models.

Everyone’s waiting on movement at the model layer. Most organisations have a bigger problem upstream: data and engineering. Fix that first, and the model becomes the easy part.

01 / Inconsistent data formats

Your data lives in five systems with three schemas and two date formats. Models trained on this either refuse to converge or learn the wrong things.

02 / Missing and duplicate records

Customer X exists three times under slightly different spellings, with key fields missing on each. Your “ground truth” is fiction without dedup and validation.

03 / No governance, no lineage

You can’t explain where the training data came from, what was excluded, or which version produced last month’s metric. Audit-time conversations end badly.

04 / Burning money on bad data

Tokens cost real money. Every retry on a malformed prompt, every re-train on a tainted batch, every agent loop on bad context, it adds up fast.

How we make your data AI-ready

The unglamorous work that makes the rest work.

Data audit & quality assessment

Honest read of what you have. We tell you which datasets are AI-ready, which need work, and which to leave behind.

Cleaning & transformation

Automated pipelines for dedup, validation, schema enforcement. Idempotent, replayable, versioned.

Governance & lineage

Catalog, lineage, access control. Auditable answers to “where did this number come from?”

Agentic workflows & RAG

Multi-step LLM workflows with tool-use, retrieval, and self-correction. Production patterns, not toy demos.

Evaluation harnesses

Test suites for prompts and agents. Regression catches before deploy. Real metrics, not vibes.

Your team owns the process

Documentation, runbooks, prompt registries. When we leave, your team operates and extends everything.

Our process

From messy data to production agents.

A proven five-phase methodology. Audit the data; build the pipelines; ship the agents; monitor the evals; hand over the keys.

Phase 01

Data audit and quality assessment

We map your data sources, identify quality gaps, and tell you honestly what is AI-ready and what isn’t.

Phase 02

Cleaning, governance, and pipelines

Automated pipelines for ingest, deduplication, validation, lineage. Governance baked in from day one.

Phase 03

Agentic workflows and ML

LLM agents with tool-use, RAG pipelines, evaluation harnesses, model deployment. Production architecture, not demos.

Phase 04

Monitoring and evals

Drift detection, agent step-completion metrics, eval pass-rate dashboards. Know when the system regresses before users do.

Phase 05

Handover and ownership

Documentation, runbooks, prompt registries, your accounts under your control. Your team owns and operates the system.

01
Phase 01

Data audit and quality assessment

We map your data sources, identify quality gaps, and tell you honestly what is AI-ready and what isn’t.
02
Phase 02

Cleaning, governance, and pipelines

Automated pipelines for ingest, deduplication, validation, lineage. Governance baked in from day one.
03
Phase 03

Agentic workflows and ML

LLM agents with tool-use, RAG pipelines, evaluation harnesses, model deployment. Production architecture, not demos.
04
Phase 04

Monitoring and evals

Drift detection, agent step-completion metrics, eval pass-rate dashboards. Know when the system regresses before users do.
05
Phase 05

Handover and ownership

Documentation, runbooks, prompt registries, your accounts under your control. Your team owns and operates the system.

AI, responsibly

The method doesn’t change. The shape of the AI does.

Most AI projects die in the data layer, not the model layer. So the five phases run the same way whether you’re building agents, retrieval, or a model pipeline. What changes is what gets built in phase three.

Phase 01

Audit the data

Which operational data an agent can safely act on, and which will make it lie.What is actually in the corpus, and how much of it quietly contradicts itself.Whether the training data can support the claim you want to make with it.Where in the workflow an assistant earns its place, and where it just gets in the way.

Phase 02

Ground it

Dedup, entity resolution, schema enforcement. The agent’s context stops being fiction.Normalisation and entity resolution first. Retrieval over duplicates returns confident nonsense.Versioned, replayable pipelines with lineage. Every number traceable back to a source row.The product data model an assistant can safely read from and write back to.

Phase 03

Build the system

Multi-step agentic workflows with tool use, retries, and an audit log on every action.Trigrams for the bulk of the matching, embeddings only where the problem is genuinely semantic.Training and deployment, drift detection from the first commit, fine-tuning only where retrieval genuinely cannot reach.Streaming UI, structured fallbacks, multi-provider routing. Three or four nines on commodity APIs.

Phase 04

Prove it

Step-completion metrics and a golden set the agent is not allowed to regress against.A retrieval eval set scored every release, where refusing beats answering wrongly.Regression suites on every commit and a pass-rate dashboard your team actually reads.An eval harness plus real user sessions before it goes near a general release.

Phase 05

Hand it over

Prompt registry, runbooks, your cloud accounts. Your team ships the next agent.Index rebuild jobs, the eval suite, and the docs your team needs to extend both.Model registry, retraining runbook, and the governance record an audit will ask for.Your team ships the next AI feature without needing to call us.

01
Phase 01

Audit the data

Which operational data an agent can safely act on, and which will make it lie.What is actually in the corpus, and how much of it quietly contradicts itself.Whether the training data can support the claim you want to make with it.Where in the workflow an assistant earns its place, and where it just gets in the way.
02
Phase 02

Ground it

Dedup, entity resolution, schema enforcement. The agent’s context stops being fiction.Normalisation and entity resolution first. Retrieval over duplicates returns confident nonsense.Versioned, replayable pipelines with lineage. Every number traceable back to a source row.The product data model an assistant can safely read from and write back to.
03
Phase 03

Build the system

Multi-step agentic workflows with tool use, retries, and an audit log on every action.Trigrams for the bulk of the matching, embeddings only where the problem is genuinely semantic.Training and deployment, drift detection from the first commit, fine-tuning only where retrieval genuinely cannot reach.Streaming UI, structured fallbacks, multi-provider routing. Three or four nines on commodity APIs.
04
Phase 04

Prove it

Step-completion metrics and a golden set the agent is not allowed to regress against.A retrieval eval set scored every release, where refusing beats answering wrongly.Regression suites on every commit and a pass-rate dashboard your team actually reads.An eval harness plus real user sessions before it goes near a general release.
05
Phase 05

Hand it over

Prompt registry, runbooks, your cloud accounts. Your team ships the next agent.Index rebuild jobs, the eval suite, and the docs your team needs to extend both.Model registry, retraining runbook, and the governance record an audit will ask for.Your team ships the next AI feature without needing to call us.

Agentic workflows
RAG systems
Model governance
Evaluation harnesses
Fine-tuning
Data preparation

What you actually get

Real AI infrastructure, not slides.

90%+

Eval pass-rate before deploy

Regression suites run on every commit. We don’t ship agents that regress against your golden set.

8, 12 wks

From data audit to production

Typical engagement timeline from first conversation to a deployed agent serving real traffic.

100%

Ownership at handover

Code, prompts, eval suites, infra, runbooks. Your accounts, your control, your IP.

Recent outcomes

Data and AI work that shipped.

Princeton University

Research-Grade Data Governance

IRB-grade governance, lineage, and tiered access across Telegram, Twitter, news, radio, TV. The kind of provenance every AI training set should have and almost none do.

3,000+ TAM · 100+ active users

Read case study

Global compliance advisory

Entity Resolution in Compliance

Financial-crime detection platform with entity resolution across messy real-world records. Higher accuracy, faster processing, deployed under heavy regulation.

Regulated · 4-month delivery

Read case study

NASA / JPL

PIXLISE Pipeline

Streaming data pipeline for Mars rover instrument data, the kind of unforgiving real-time data work that AI agents claim to handle and almost never do. NASA Software of the Year runner-up.

30hrs → 10min

Read case study

Resources for AI buyers.

Self-assessment

The AI-Readiness Audit

A practical six-dimension audit for telling whether a data foundation can carry AI, before the budget is committed. About ten minutes, one honest conversation.

Run the audit Reliability playbook

How to Improve LLM Uptime

Multi-provider routing, hedged requests, structured fallbacks, model version pinning, and the SLOs to track. The playbook we use to get LLM-backed products to three or four nines on commodity APIs.

Read the playbook

Heard on the AI Briefing podcast

Episode

LLM Uptime Crisis: What happens when AI services go offline

Episode

AI Implementation Strategy: Why data fundamentals still matter in the age of LLMs

Episode

Building on AI: How much risk can you handle?

Questions buyers ask first

The honest answers to what you're about to ask.

Are you an AI consultancy, or a data engineering consultancy doing AI?: The second. The honest read is that most “AI projects” are 80% data engineering and 20% model work, and most “AI consultancies” are PowerPoint shops that subcontract the data engineering. We've been doing the data engineering underneath production systems for 15 years; AI just changed what the output layer is called.
Will you fine-tune a foundation model for us?: Almost never, and we'll usually talk you out of it. Most use cases that founders think need fine-tuning are actually retrieval problems, prompt design problems, or data quality problems. We'll tell you in the audit whether yours genuinely needs fine-tuning; if so, we'll partner with a model-shop or run it ourselves depending on the shape.
How do you handle RAG and retrieval over messy data?: Entity resolution and normalisation before anything else. A retrieval system over duplicate, inconsistent records gives the model confidently-wrong answers, which is worse than refusing. We layer pg_trgm for the bulk of the matching, Jaro-Winkler for short identifiers, and vector embeddings only where the problem is genuinely semantic.
What about evals and regression testing for agents?: Every engagement ships with an evaluation harness, a golden test set, regression suites, and a pass-rate dashboard. We don't deploy agents that regress against the golden set. “It felt better in vibes” is not a release criterion.
Can you operate in regulated environments (finance, health, compliance)?: Yes. We built a sanctions-compliance platform for a global financial-crime advisory that runs in a heavily-regulated environment across 20+ financial institutions; the Princeton Research Accelerator built IRB-compliant research data governance. AI work in regulated industries is mostly about lineage, audit trails, and access control, which is the unsexy data engineering we lead with anyway.
How long does an AI engagement typically take?: Eight to twelve weeks for data audit through deployed agent serving real traffic, the standard shape. Longer if the audit reveals that the data work has to come first (which it often does, and we’d rather you know that early). Phased and milestone-bound; you can stop at any phase boundary.

For enterprise

AI in regulated industries

Governance, audit trails, security posture. AI work that survives procurement and compliance review.

Visit /for-enterprise

For startups

Agentic features for your product

LLM-backed features in your MVP. Production architecture from sprint one, not a demo that breaks at scale.

Visit /for-startups

Send Tom the agent you’re trying to make work.

A few lines about the data, the prompt, or the agent that won’t hold up under real traffic. He’ll write back honestly about where to dig in first.

Talk to our team