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
Retrieval-augmented generation
Combine a generator with retrieved source material, then evaluate retrieval quality and answer grounding separately.
Lewis et al., RAG paper, opens in a new tab
Model adaptationWays to specialize model behavior without retraining every parameter.1 entries
Low-rank adaptation
Train small low-rank matrices while keeping the pretrained model weights fixed, when the task and evidence justify adaptation.
Hu et al., LoRA paper, opens in a new tab
Tool and context interfacesProtocol-level boundaries for connecting AI applications to tools and data.1 entries
Model Context Protocol
Use an explicit protocol boundary for tools, resources, and prompts, with authorization and deployment decisions treated as part of the system design.
MCP 2026-07-28 specification release, opens in a new tab
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
Reasoning and acting loops
Interleave reasoning traces with task-specific actions, while keeping tool authority, failure handling, and evaluation explicit.
Yao et al., ReAct paper, opens in a new tab
Attention kernelsImplementation techniques that reduce data movement for transformer attention.1 entries
FlashAttention
Use an input-output-aware exact attention algorithm when supported hardware and workload measurements show a practical benefit.
Dao et al., FlashAttention paper, opens in a new tab
Model servingMemory management patterns for higher-throughput language model inference.1 entries
PagedAttention and vLLM
Manage key-value cache memory in blocks so serving decisions can account for fragmentation, batching, and changing sequence lengths.
Kwon et al., PagedAttention paper, opens in a new tab
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
GraphRAG
Evaluate graph-derived summaries for global or multi-part questions, including the additional indexing cost and evidence requirements.
Microsoft Research, GraphRAG publications, opens in a new tab
Multimodal retrievalRetrieval systems that select evidence across text and visual inputs.1 entries
Self-adaptive multimodal RAG
Track research that selects retrieval strategies across modalities, then validate it against the actual document, image, and latency constraints of the use case.
Yu et al., multimodal RAG paper, opens in a new tab
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.
| Category | Production-ready | Under evaluation | Research watchlist |
|---|---|---|---|
| Primary question | Can 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 expected | Task-level quality, cost, latency, security, and failure behavior | Prototype results against a simpler baseline | Paper findings, reproducible code, and a narrow hypothesis |
| Delivery posture | Production candidate after system-specific validation | Controlled evaluation before production use | Research watchlist, not a default recommendation |