Skip to content
Back to Surviant

Surviant engineering reference / AI + ML

AI and machine learning capability reference.

Purpose
Support product and architecture planning
Scope
Methods relevant to current AI and ML work
Status
Working reference
Reviewed
August 28, 2026

Classification method

Readiness describes the maturity of the method and its supporting ecosystem. Selection still depends on data, security, latency, cost, maintainability, and measurable user value.

Maturity / Production

Production-ready methods

Methods with stable tooling, understood failure modes, and established deployment patterns. Inclusion does not mean every project uses them.

Knowledge groundingPatterns for connecting generation to selected external material.1 entries
Model adaptationWays to specialize model behavior without retraining every parameter.1 entries
Tool and context interfacesProtocol-level boundaries for connecting AI applications to tools and data.1 entries

Maturity / Evaluating

Methods under evaluation

Methods with credible implementations and growing adoption. We validate them through bounded prototypes before recommending production use.

Agent patternsModel-led task execution that alternates between reasoning and external actions.1 entries
Attention kernelsImplementation techniques that reduce data movement for transformer attention.1 entries
Model servingMemory management patterns for higher-throughput language model inference.1 entries

Maturity / Watching

Research watchlist

Early methods tracked through papers, reference implementations, and small experiments. These are not presented as client-ready capabilities.

Sequence architecturesAlternatives to attention-based sequence modeling that need workload-specific validation.1 entries
  • Mamba and selective state spaces

    Explore input-dependent state-space models for sequence workloads where linear scaling may matter, without assuming a general replacement for transformers.

    Gu and Dao, Mamba paper, opens in a new tab
Graph-based retrievalStructured retrieval for questions that depend on relationships across a corpus.1 entries
Multimodal retrievalRetrieval systems that select evidence across text and visual inputs.1 entries

Reference / At a glance

Compare the maturity bands.

This table is a navigation aid for product conversations. Readiness still depends on the use case, data, risk, operating cost, and quality threshold.

Scroll horizontally to compare columns.

Comparison of production, evaluating, and watching approaches by AI and machine learning category
CategoryProduction-readyUnder evaluationResearch watchlist
Primary questionCan this established pattern solve the measured task?Does a bounded prototype justify operational adoption?Is this worth monitoring or testing in a research setting?
Evidence expectedTask-level quality, cost, latency, security, and failure behaviorPrototype results against a simpler baselinePaper findings, reproducible code, and a narrow hypothesis
Delivery postureProduction candidate after system-specific validationControlled evaluation before production useResearch watchlist, not a default recommendation