Stack 01 · AI & Machine Learning
A production AI architecture should make model choice important but replaceable. Aviox separates model access, retrieval, business logic, evaluation and monitoring so one vendor does not become the entire application.
How we choose
Technology is chosen after the operating constraint: security, residency, latency, quality, cost and integration requirements drive the stack.
The fastest way to regret an AI architecture is to build it around one model provider's API shape. Prices move, context windows triple, a better model ships from somewhere else, and suddenly the thing you cannot change is the thing you most want to. So we put an abstraction between the application and the model from day one — not an elaborate one, just enough that swapping a provider is a configuration change and a re-run of the evaluation suite.
The same discipline applies to retrieval and training. Embeddings and indexes are rebuildable from source. Fine-tuned models are reproducible from versioned datasets. Nothing in the pipeline is a one-off artefact that only one person knows how to recreate.
Non-negotiables
The toolchain
Each row says what the tool is used for and links to related pages. Named tools are examples of what can sit in each layer, not a list of partnerships or certifications.
Our default for extraction, classification and generation where the data can leave the boundary. Used behind an abstraction layer so a model or provider change is configuration, not a rewrite.
Where a whole claim pack, disclosure bundle or policy document has to be reasoned over in one pass, and where tool-using agents need to stay on task across many steps.
Agent graphs with explicit state, tool calling and human-in-the-loop checkpoints — so an agent's path through your systems is a thing you can draw, test and replay.
Document ingestion, chunking strategies and hybrid retrieval. This is where most of the quality in a RAG system actually comes from, so it is where most of our tuning time goes.
Custom scoring and ranking models, and fine-tunes where a hosted model cannot reach the accuracy a rubric demands. Reproducible from versioned datasets, not from someone's laptop.
Plenty of clients already have TensorFlow in production. We work with what is there rather than insisting on a rewrite, and bring it under the same CI, registry and monitoring as everything else.
Open-weight models self-hosted where data residency, cost at volume or air-gapped deployment rules out a hosted API. Same evaluation gate, different inference location.
Embeddings stored next to the relational data they describe, so permission scoping and joins are trivial. Dedicated vector stores only when scale genuinely demands one.
Automatic speech recognition and pronunciation scoring — the layer that makes instant speaking assessment possible, and that powers call transcription and quality work elsewhere.
Golden datasets, faithfulness and refusal scoring, agreement checks against human experts, and red-team suites — all running in CI. Nothing with a model in it ships without passing.
Built on this stack
The useful question about a technology list is what has been built with it. These are the Aviox-operated product and the published case study that apply this stack.
An AI-assisted practice platform that Aviox builds and operates, applying speech and language models, criterion-level scoring and evaluation to practice attempts.
The insurance workflow documented in the case study: document extraction, retrieval, rules and human review, with the source of every field shown to the reviewer.
Scoring and feedback workflows built around an institution's own criteria, with variance tested on representative submissions before release.
Note Products we run ourselves patterns are starting points configured to your data and rules Read the case study
Request a quote
Talk to an engineer about your architecture, or get a quote for the work.
In the field
Three sectors where this particular layer carries the most weight. Each page shows the five-step automation pattern in that sector's own vocabulary.
Claims intake, extraction and triage with evidence-backed work queues.
Explore Industry 07Disclosure review and contract analysis, cited down to the clause.
Explore Industry 08Assessment and feedback workflows, with an Aviox-operated product as a reference.
ExploreAlso All ten industries run on some combination of these four stacks
Connected
No system we build uses only one of these four layers. Here is the service line that leads the work, and the stacks that sit either side of this one.
Services that use this stack
Retrieval grounded in your own records, evaluation harnesses and the guardrails that keep answers on-policy.
Explore Service 02Agents that triage, update records and call your systems — escalating only the exceptions, every step logged.
Explore Service 04Forecasting, propensity and risk models with feature stores, drift monitoring and explainability built in.
ExploreThe other stacks
Questions
Asked most often by engineering and product teams.
We pick for the problem and keep the choice swappable. Retrieval, business logic and evaluation are designed so one model or provider does not become the architecture.
With a representative test set that includes difficult and adversarial cases, with retrieval evaluated separately from the final task, and re-run whenever models, prompts, tools or source data change.
Often, yes. Hosting, model provider and data flows are agreed during scoping against your security rules.
Through logs, quality sampling, cost tracking and alerts, with a fallback to a person when confidence is low.
Request a quote
Share what you are building and the constraints around it. We will say where model choice matters, what to measure and what to keep replaceable.
