Skip to content
7 Hidden LayersBeyond the API
AI Engineering & Research Consultancy

Depth is thedifference.

Anyone can call an API. We work the ten layers underneath it — where latency, cost and ownership are actually decided.

Book a consultationDescend the stack↗↗
↓Scroll to descend

The lower layers decide whether a system is fast, affordable, and yours.

RAG◆Multi-agent◆Fine-tuning◆Quantisation◆Distillation◆Inference◆Evaluation◆On-prem◆GPU◆Serving◆Small models◆Research◆RAG◆Multi-agent◆Fine-tuning◆Quantisation◆Distillation◆Inference◆Evaluation◆On-prem◆GPU◆Serving◆Small models◆Research◆

Most AI work stops at the surface. A prompt, an API key, a demo that impresses in a meeting and buckles in production. We go further down — because the layers nobody shows you are the ones that decide whether a system is fast, affordable, and actually yours.

01

Depth over demos

A demo proves a model can. Depth proves a system will — at your latency, your cost, your load.

02

Own the lower layers

The layers nobody shows you decide whether a system is fast, affordable, and actually yours.

03

Measure or don't claim

Every improvement we report has an evaluation behind it that you can run yourself.

04

Leave nothing undocumented

We are consultants. The engagement ends. The system should not notice.

[02The stack]

We work the whole stack.

Ten layers between a question and an answer. Most consultancies rent you the top one. Open any layer to see what we actually do down there.

  • The part everyone sees, and the only part most vendors touch. Interfaces that make a probabilistic system feel dependable — streaming, citations, graceful failure, and the affordances that let a person stay in control.

    • Streaming UX
    • Human-in-the-loop
    • Trust & citations
    • Latency budgets
Explore the full stack↗↗Ten layers, one system
[01Capabilities]

Six ways in.

Engagements are shaped around the layer your problem actually lives in — not around a package we happen to sell.

01Layers 01–04

AI Strategy

Where to apply AI, what to build, and what to buy.

  • /Opportunity mapping
  • /Build vs. buy analysis
  • /Model & vendor selection
  • /Infrastructure economics
02Layers 01–03

AI Systems Engineering

End-to-end design and construction of production AI systems.

  • /RAG architecture
  • /Multi-agent systems
  • /Evaluation frameworks
  • /Observability
03Layers 05–06

Model Optimization

Fine-tuning, distillation, and small language models.

  • /LoRA & preference tuning
  • /Distillation
  • /Small language models
  • /Task-specific evals
04Layers 07–10

Inference Infrastructure

Serving stacks engineered for throughput and cost.

  • /Quantisation
  • /Continuous batching
  • /GPU deployment
  • /Cost per token
05Layers 04–10

Enterprise AI

AI that survives procurement, security review, and scale.

  • /On-prem & VPC
  • /Data residency
  • /Security review support
  • /Audit trails
06Any layer

Research & Prototyping

When the answer isn't in a paper yet.

  • /Applied research sprints
  • /Feasibility spikes
  • /De-risking prototypes
  • /Written findings
All six capabilities, in detail→
[03Approach]

How an engagement runs.

  1. Step 01

    Discovery

    We start with the constraint, not the technology. What must be true for this to be worth doing — and what happens if it isn't.

    Problem brief & success criteria
  2. Step 02

    Architecture

    The system drawn before it is built. Layer by layer, with the trade-offs written down where they can be argued with.

    Architecture & decision record
  3. Step 03

    Prototype

    The riskiest assumption, built first and measured against a real evaluation set. Fast enough to be wrong cheaply.

    Working prototype & eval harness
  4. Step 04

    Optimization

    Down the stack. Quality, latency and cost pushed until the curve flattens and further effort stops paying.

    Benchmarked performance envelope
  5. Step 05

    Production Deployment

    Shipped with the unglamorous parts intact: monitoring, rollback, on-call notes and a team that can operate it without us.

    Live system & runbook
  6. Step 06

    Continuous Improvement

    Models move, data drifts, costs change. A cadence of evaluation and tuning that keeps the system from quietly decaying.

    Ongoing eval & tuning cadence
[04Work]

Depth, in practice.

Three engagements, anonymised. Detailed case studies are available on request.

01Financial Services

Retrieval that survives an auditor

A research desk had a document assistant that answered fluently and cited badly. Confidence was high, grounding was not, and compliance had stopped the rollout.

Answer groundedness
3.1×
Answer groundedness
Unsupported claims
−62%
Unsupported claims
Compliance review
Passed
Compliance review
02Consumer Platform

The bill was in the seventh layer

Inference spend was growing faster than usage. The product team had optimised prompts twice and run out of surface-level ideas.

Cost per request
−71%
Cost per request
Throughput
2.4×
Throughput
Quality parity
Held
Quality parity
03Healthcare

An assistant that never leaves the building

Clinical data could not cross the network boundary, which ruled out every hosted option the team had evaluated.

Data egress
0
Data egress
Idle GPU reclaimed
40%
Idle GPU reclaimed
Deployment
On-prem
Deployment
[05By the numbers]
0

Layers of the stack, worked end to end

≤ 0wk

From first call to a measurable prototype

0×

Typical inference cost reduction

0%

Engagements handed over with a runbook

Next step

Let's build AI systems that last.

A first conversation costs nothing and usually saves a quarter. Bring the constraint you're stuck on — we'll tell you which layer it lives in.

Book a consultationEmail us↗↗

Typical reply within one working day.