Technologies · AI Models

The AI models we build with — and how we pick them.

We are model-agnostic by design. Every build starts with the cheapest capable model that clears the quality bar, and scales up only where the task demands it. Grounded, evaluated, and never locked to one vendor.

Least-cost model laddercheapest that clears the bar
L0
Local & open weights
$0 floor
L1
Cost-efficient hosted
high volume
L2
Mid-tier reasoning
user-facing
L3
Strong models at the gate
hard reasoning
Why it matters

Most teams over-pay for the wrong model.

The model is the smallest part of a working AI product — but it is where budgets quietly bleed. We treat model selection as an engineering decision, not a default.

Picking a model is a cost-versus-quality trade-off made per task, not per project. A classification step, a summarisation step, and a customer-facing reasoning step have completely different requirements — so we route each to the right tier. The result is production AI that is grounded, measured against real evaluation sets, and typically a fraction of the cost of a single-model build. Across our own fleet of 33 built products, this discipline is why 18 run live in production without a runaway inference bill.

0
model families in rotation
0
products built on the ladder
0+
agents in production
0
live platforms
The least-cost model ladder

Cheapest capable model that clears the bar wins.

We climb the ladder only when the task needs it. $0 local and open models first; a strong paid model only at the quality gate.

L0
Local & open modelsLlama and other open weights, self-hosted for classification, extraction, and routine generation.
$0 floor
L1
Cost-efficient hosted modelsFast, inexpensive hosted models — DeepSeek, Mistral, and small-tier models — for high-volume work.
Low cost
L2
Mid-tier reasoningBalanced hosted models for grounded generation, agents, and RAG answers that face users.
Metered
L3
Strong models at the gateTop-tier reasoning from OpenAI, Anthropic, or Gemini — reserved for hard reasoning and final-quality gates.
Gated
Model families in rotationswappable layer
OpenAI · Anthropic
final-gate
Gemini
long-context
Llama / open
self-hosted
DeepSeek · Mistral
volume

The families we build with.

We keep several model families in active rotation so we can match capability to task and never depend on a single provider’s pricing, availability, or roadmap.

  • OpenAI & Anthropic for hard reasoning and final-gate quality
  • Google Gemini for long-context and multimodal work
  • Llama and other open weights for the $0 self-hosted floor
  • DeepSeek and Mistral for cost-efficient, high-volume generation
  • A swappable model layer so no build is locked to one vendor

When and why we pick each.

Selection is evidence-based. We benchmark candidate models on your task with real evaluation sets, score them on cost, quality, and latency, and let the numbers pick the winner.

  • Per-task benchmarking on representative data, not vibes
  • Cost × quality × latency scoring for every candidate
  • Spend caps and circuit breakers so a paid model can never run away
  • Outputs grounded in your data and checked against ground truth
  • Continuous re-evaluation as new models ship
Per-task evaluationcost × quality × latency
Benchmark
real eval sets
Score
3 axes
Pick winner
evidence-based
FAQ

Questions we get asked

There is no single best model — the right answer is usually several. We route each task to the cheapest model that clears its quality bar, so a product may use an open model for extraction and a strong model only for its hardest reasoning step.
Yes. Open weights such as Llama are the $0 floor of our ladder and handle a large share of production work. We escalate to paid hosted models only where the task genuinely needs it.
Every paid model call runs under a spend cap with a circuit breaker. We benchmark cost against quality per task and default to the least-cost option, which is how we keep 18 live products economical.
No. The model layer is designed to be swapped. If pricing, quality, or availability changes, we move the workload without rewriting your product.
Our stack

The technology behind this.

The real, relevant stack we build this with — model-agnostic, open-source-first, production-grade.

Keep exploring

Related capabilities

Every technology below is delivered by the same composed engineering team.

Ready when you are

Have an AI feature in mind?

Tell us the outcome. We’ll pick the model ladder, ground it in your data, and ship it gated end to end.

Start a project →