Start a conversation

Independent LLM infrastructure practice

Ship language models that hold under load.

Harness Consult is a consulting practice for teams putting language models into production — in the cloud, on the device, or inside a boundary the data is not allowed to leave.

Start a conversationSee the services
p99 latency against concurrent requestsAn illustrative line showing p99 latency staying near 200 to 250 milliseconds, well under a 400 millisecond budget, as concurrency rises from 0 to 512 requests.400 ms budgetp9901282563845120200400600 msconcurrent requests →
Illustrative target shape — the curve we engineer toward, not a client benchmark.

What we do

Six categories of work.

Most engagements start in one and touch two. They are separated here because the constraints are genuinely different — a memory ceiling on a phone and a regulator's boundary are not the same problem wearing different clothes.

Black fibre-optic patch cables arcing across a dark patch panel, each port lit by a small amber indicator.

LLM cloud

Model selection, gateways, routing and cost control across the major providers and open-weight models.

  • Multi-cloud
  • Gateways
  • Cost control
A bare silicon chip on matte black anodised aluminium, its edge catching a thin rim of amber light.

On-device & edge

Phones, laptops, vehicles and embedded targets — quantisation, runtimes, memory budgets and hybrid fallback.

  • Mobile & edge
  • Quantisation
  • NPU
Four machined metal blocks of decreasing size on matte black, raking amber light revealing the tooling marks.

Optimisation

Quantisation, distillation, adapters, speculative decoding and caching — each measured against a quality bar.

  • Quantisation
  • Distillation
  • Evals
A black server blade half-withdrawn from its rack chassis on precision rails, amber light spilling from the open bay.

Deployment & serving

Serving stacks, GPU scheduling, model registries and release mechanics that survive a bad Friday.

  • vLLM / Triton
  • Kubernetes
  • Canary
A long row of identical black rack cabinets receding down a dark aisle, each marked by one amber status light.

Scaling & reliability

Capacity planning, autoscaling, queueing, SLOs and failover for workloads that cannot fall over.

  • Autoscaling
  • SLOs
  • Capacity
A closed matte black steel cage door with a precision lock, dark server racks glowing amber behind its mesh.

Private & sovereign cloud

On-prem, VPC-isolated and air-gapped deployments, with the compliance evidence that makes them signable.

  • On-prem
  • Air-gapped
  • Compliance

How we work

Measured, then built, then handed over.

Four stages, in order, because each one's output is the next one's input. The first is always measurement — a recommendation without a baseline is a preference.

Read the full process
  1. 01

    Diagnostic

    1–2 weeks

  2. 02

    Architecture

    1–3 weeks

  3. 03

    Build

    4–12 weeks

  4. 04

    Handover

    1–2 weeks

How we operate

Four commitments.

  • Measure first

    No recommendation before a baseline. Teams routinely optimise the wrong thing for a quarter because nobody put a number on it in week one.

  • Your repo, your cloud

    Work lands in your infrastructure, your IaC and your review process from day one. Nothing of consequence should live on a consultant's laptop.

  • The boring answer, when it's right

    Frequently the correct advice is a cache, a shorter prompt, or staying on the API you already pay for. We would rather be useful than interesting.

  • Leave the team stronger

    Success is your engineers operating this confidently without us. We are a temporary function by design.

Tell us what's breaking.

Bring a workload, a latency target, or a bill you can't explain. First conversation is a technical one — no discovery deck.

Start a conversation