AI Architecture Advisor

AI Architecture Advisor

The front-door decision skill for "I have problem X — what should I build with?" It owns the choice — which approach fits, when to promote complexity (and when not to), how to transfer knowledge, how to scale — then hands off to the deep skill that owns the depth. It spans the whole modeling space in one place so you can compare options that normally live in separate skills side by side:

tabular GBDT  ·  deep net  ·  Transformer/LLM  ·  RAG  ·  fine-tuning  ·  agents
multimodal/omni  ·  embeddings & retrieval  ·  recsys/ranking  ·  model architecture (dense/MoE/SSM/diffusion)

No theory dumps — decision tables, elimination logic, tradeoffs, and a pointer to the deep skill.

The architect's move is to ask before answering. The amateur hears "build an AI feature" and reaches for the model they know ("we'll fine-tune Kimi"). The architect first asks: what data type? what volume? what task? what's the success metric? do you even need a Transformer? The skill that distinguishes an architect is the willingness to say "for this, CatBoost wins," "here you need a Transformer," or "LoRA is enough here" — and to refuse to name an approach until the problem is classified. Never jump to a model before the Intake questions below are answered.

ASCII Flow

problem + data + constraints
  |
  v
0. INTAKE — ask before answering (see questions below)
  |          do NOT name a model until task + data + metric + constraints are known
  v
1. classify the problem          (tabular? text? generation? decision/action? retrieval?)
  |
  v
2. eliminate ineligible options  (with a reason each — never silently drop)
  |
  v
3. score survivors independently (accuracy, latency, cost, data need, interpretability, ops)
  |
  v
4. pick the SIMPLEST that clears the bar   (start simple, promote only on evidence)
  |
  v
5. hand off to the deep skill    (ai-ml-data-science / ai-llm / ai-rag / ai-agents ...)

Intake: Ask Before You Answer

A request like "let's fine-tune model X" is a proposed solution, not a problem statement. Do not accept it at face value. Surface the six questions that decide the architecture, and hold any model name until they're answered. If the user can't answer one, that gap is itself a finding (most often: no success metric, or no labeled data).

Breaking ties among survivors: hard constraints (a p99 ceiling, an on-prem requirement, a regulatory bar) are gates — fail one and the option is out, no matter how it scores elsewhere. Only among options that clear every gate do the soft criteria (cost, developer familiarity, marginal accuracy) act as tie-breakers. Don't let a soft criterion overrule a failed hard constraint, and don't eliminate an option on a soft criterion before checking whether a rival actually clears every gate.

Quick Reference

When to Use This Skill

Activate when the user asks (in any language) some form of:

• "What architecture should I use for X?" / "Как выбрать архитектуру?"
• "When should I use CatBoost / Transformer / RAG / fine-tuning / an agent?"
• "Trees or neural net for this tabular problem?"
• "Do I need RAG, or is fine-tuning better?"
• "Should this be one LLM call or a multi-agent system?"
• "Omni model or a specialist pipeline for my images/audio?"
• "Which embedding model / do I need a reranker?"
• "Is this a search problem or a recommendation problem?"
• "Dense or MoE / should I care about Mamba / diffusion LLMs?" (model-architecture axis)
• "How do I find and evaluate newer architectures than what I know?"
• "How do I transfer knowledge to a new task?" (transfer learning / distillation / RAG)
• "How do I scale this model?" (data, params, test-time compute, serving, or distributed training)
• Any side-by-side comparison of modeling approaches before committing to a build.

If the user has already chosen the approach and wants to build it, skip this skill and go straight to the deep skill.

Scope Boundaries (Use These Skills for Depth)

This skill decides; these implement. Hand off once the approach is chosen. Primary skill per lane below — the full AI skill map catalogs all 36 ai-* skills (supporting + adjacent) grouped by lane, with one-line "owns what" descriptions.

• Classical ML / data → ai-ml-data-science (GBDT, EDA, train a specialist), ai-ml-timeseries
• LLM lifecycle & prompting → ai-llm (fine-tune, migrate, select), ai-prompt-engineering, ai-post-training (RLHF/DPO/GRPO/RLVR)
• Retrieval & context → ai-rag, ai-context-layer, ai-vector-brain
• Agents & applied bots → ai-agents, ai-bot-builder, ai-voice-bots, ai-coding-agents
• Pre-training → ai-pretraining, ai-distributed-training, ai-scaling-laws, ai-data-curation-pretraining
• Serving, evals & ops → ai-llm-inference, ai-local-model-ops, ai-evals, ai-deep-research, ai-mlops, ai-product-operating-model
• Scouting new architectures → research-arxiv-scout, research-scout
• Beyond the AI family → foundations-decision-theory (choice under uncertainty), software-architecture-design (system/runtime topology)

Default Workflow

1. Classify the problem. What is the output and the data?
   • Structured/tabular rows -> classical ML lane.
   • Natural-language understanding or generation -> LLM lane.
   • Retrieve-then-answer over a corpus -> RAG lane.
   • Multi-step actions / tool use -> agent lane.
   • Images/audio/raw signals -> deep-learning lane.
2. Eliminate ineligible options with a reason. Never silently drop a candidate. E.g. "fine-tuning eliminated: the gap is missing knowledge, not behavior — RAG first."
3. Score the survivors independently on: accuracy ceiling, latency, cost-per-outcome, data required, interpretability, and operational burden. Don't pre-commit before scoring.
4. Pick the simplest option that clears the bar. Start simple; promote complexity only on evidence of a repeated, stable failure the simpler option can't fix.
5. State the decision and hand off to the deep skill, with the reason and the promotion trigger ("revisit if X").

Decision Tree: The Lanes

What are you predicting / producing?
    |
    +- A value/label from columns (tabular) ............ CLASSICAL ML
    |     small data or need interpretability? ............ trees / linear
    |     within a tabular-FM envelope, max accuracy? ..... tabular foundation model (TabPFN-3 / TabICL v2)
    |     accuracy on large/wide structured data? ......... gradient-boosted trees
    |
    +- Rank items for a user from behavior ............. RECSYS / RANKING
    |     implicit feedback + <50ms + dynamic catalog? ... two-tower retrieve -> ranker (not RAG)
    |
    +- Text understanding/generation .................... LLM
    |     stable instructions, no private data? .......... prompt-only
    |     multi-step logic/planning gap? ................. raise thinking budget (dial, not a separate model)
    |     needs current/private knowledge? ............... + RAG / context engineering
    |     stable behavior/format gap remains? ............ + fine-tune (SFT)
    |     preference/safety/reasoning gap after SFT? ..... + post-train (RLHF/DPO/GRPO/RLVR) -> ai-post-training
    |
    +- Answer grounded in a document corpus ............. RAG (pick the embedder + reranker first)
    |     relational/multi-hop over many entities? ...... GraphRAG / agentic RAG
    |
    +- Take actions / call tools / multi-step .......... AGENT
    |     one tool, linear? .............................. tool-use workflow
    |     planning + many tools? ......................... single agent (tools via MCP)
    |     distinct specialized + parallel roles? ........ multi-agent (only if single provably can't)
    |
    +- Images / audio / video + text ................... MULTIMODAL
    |     validating a product? .......................... omni / VLM model
    |     accuracy / cost-at-volume / no-hallucination? .. specialist per modality -> LLM reasoning
    |
    +- Images / audio / raw signals (no language) ...... DEEP LEARNING
    |     spatial grid (images)? ......................... CNN / Vision Transformer (ViT)
    |     ordered sequence, tiny/streaming footprint? .... RNN / LSTM / GRU (else Transformer)
    |     text-to-text transform (translate/summarize)? .. encoder-decoder (T5-class seq2seq)
    |     encode text for classification/retrieval? ...... encoder-only (BERT-class)
    |     generate images/audio? ......................... diffusion / GAN / VAE (see emerging note)

Orthogonal to the lane — only when self-hosting or training the weights:
    model architecture = dense | MoE | SSM/hybrid | diffusion-LM   (invisible behind an API)
    attention variant   = MHA -> GQA (default) | MLA (KV-cache) | NSA/sliding-window (long-context)

Older/foundational nets are still the right call sometimes — name and route them, never silently drop. RNN/LSTM/GRU for tiny or strictly-streaming sequence models; CNN for spatial image features; encoder-only (BERT-class) as the workhorse embedding/classification backbone behind RAG and rankers; encoder-decoder (T5-class) for fixed text-to-text transforms. Decoder-only LLMs dominate open-ended generation, but these are not obsolete — they win on footprint, latency, and task fit. Depth handoffs: ai-ml-data-science (CNN/RNN/training a specialist), ai-rag + ai-vector-brain (encoder-only embedders).

Classical ML: Which Tree Library

For tabular data, gradient-boosted trees are the strong default. The three libraries are the same algorithm with different engineering — pick by the deciding feature, not hype.

Rule of thumb: start RandomForest (baseline) -> XGBoost (tune) -> switch to LightGBM if too slow on big data, or CatBoost if categorical-heavy. Reach for deep learning on tabular only when raw images/text/audio are part of the row.

2026 update — tabular foundation models: "trees always win on tabular" (Grinsztajn 2022) no longer holds within an in-context tabular-FM's envelope (TabPFN-3 ~1M rows; TabICL v2 if the non-commercial TabPFN license blocks you). Try a tabular FM first when the table is within its version-specific envelope and raw accuracy is the bar; trees stay the default for very wide tables, sizes beyond the envelope, and where a natively inspectable, CPU-cheap model matters. The full lineage + signal-by-signal GBDT-vs-FM table (sizes are volatile — verify before quoting) is in decision-matrices.md.

LLM Lane: Prompt -> Reasoning -> RAG -> Fine-tune -> Agent

Promote one rung only when the current rung provably fails. Match the rung to the gap type — reasoning, knowledge, and behavior are different failures with different fixes:

AWS cloud-provider fork — when "must run on AWS" is the binding constraint, the lane logic above still applies; the hosting tier just maps onto AWS services. Inference: Bedrock (managed, multi-provider, zero-ops) → SageMaker (custom/fine-tune serving) → EC2/EKS (full OSS self-host). Managed RAG by ownership scope: Bedrock Knowledge Bases (zero-ops) / Kendra GenAI Index (retriever reused across surfaces) / Q Business (no-code end-to-end assistant). Governance: Bedrock Guardrails (grounding check, PII, denied topics, prompt-attack — model-independent via ApplyGuardrail). Full service-by-service mapping in decision-matrices.md.

(Detailed RAG-type and model-tier matrices live in ai-llm/references/decision-matrices.md.)

Model Architecture: Dense vs MoE vs SSM/Hybrid vs Diffusion

This axis is orthogonal to the lane and invisible behind an API — it only becomes a decision when you self-host weights or train from scratch. On an API, ignore this entire axis (the knobs that matter there are model tier and thinking budget). Self-hosting → start dense unless a binding constraint forces otherwise: MoE (cost-per-token at scale — the frontier default), SSM/hybrid (>100k-tok throughput; only the hybrid is competitive — pure SSM fails in-context learning), or diffusion-LM (raw tokens/sec for code/structured text). Then pick the MoE granularity and attention variant against that constraint. The full architecture-comparison + sub-axes tables are in decision-matrices.md; hand off to ai-llm-inference (serving) and ai-distributed-training (training).

"What limits this component and how do I work around it?" The advisor routes, it doesn't carry the mechanics. The Architecture Limitation → Workaround → Owning Skill table (attention O(n²), KV-cache, RoPE extrapolation, MoE cost, quantization outliers, …) and the Encoder vs Decoder vs Encoder-Decoder family table map each one; build-time depth in ai-pretraining, serve-time in ai-llm-inference.

Multimodal: Omni Model vs Specialist Pipeline

For images/audio/video alongside text, the choice is one unified model vs a specialist perceptual layer feeding an LLM.

• Omni / VLM (single model) — default to validate a product fast; one API, broad coverage. Good enough for general visual Q&A, light OCR, casual audio.
• Specialist per modality -> LLM reasoning — switch when accuracy is the KPI, cost at volume binds, or you need non-hallucinating outputs (medical imaging, high-volume OCR, real-time/streaming ASR). Specialists are often far cheaper and more accurate on their modality; non-autoregressive ASR is structurally hallucination-resistant — a gap that does not close with frontier model scale. Enterprises commonly run many specialist models behind a reasoning LLM rather than one omni model. OCR / IDP routing (AWS Bedrock Data Automation vs Textract standalone vs self-hosted paperless-ngx) is mapped in decision-matrices.md.

Hand off to ai-prompt-engineering (multimodal prompting) and ai-ml-data-science (training a specialist).

Embeddings & Retrieval Model (a decision with lock-in)

The embedding/retrieval model is a first-order choice, not "RAG config": swapping it forces a full re-index. Decide deliberately on five axes (modality → hosting → language → domain fit → retrieval architecture), and add a cross-encoder reranker by default (skip only under a hard p99 budget). Five-axis selection tree in decision-matrices.md; depth in ai-rag, concrete brains in ai-vector-brain.

Recommendation / Ranking (a distinct lane, not RAG)

Implicit behavioral signals (clicks/dwell/skips) + <50ms + a dynamic catalog of millions = a recommender, not RAG — even though both "retrieve then rank." Don't model it as vector-RAG; the default shape is two-tower retrieval → ranking model. RAG-vs-recsys property table in decision-matrices.md; depth in ai-ml-data-science plus RecSys literature.

How to Transfer Knowledge

"Transfer knowledge to a new task" has several distinct mechanisms — pick by what's stable and what changes: new facts → RAG (no training); new domain with labels → fine-tune/LoRA; shrink to a cheap model → distillation (reasoning distillation to keep CoT); new downstream task on a vision/audio backbone → freeze + train a head; related tabular task → retrain trees or let a tabular FM transfer via in-context learning. Situation-by-mechanism table in decision-matrices.md.

How to Scale a Model

"Scale" is ambiguous — clarify which axis first, because the answer differs: (a) better quality from more data/params (scaling laws), (b) better quality without retraining (test-time compute — now a primary lever, not an afterthought: thinking budget / best-of-N + verifier, traded against latency and cost per query class), (c) train a model too big for one GPU (FSDP/ZeRO), (d) serve more traffic / lower latency (batching, quantization, routing), (e) more data in the pipeline (data-lake/streaming). Axis → move → deep-skill routing table in decision-matrices.md.

How to Find & Vet New Architectures

The field moves faster than this skill can be re-edited. When the right approach might be newer than what's catalogued, discover and vet deliberately — don't adopt on hype or reject on unfamiliarity. The discipline in one line: scout a fixed signal set → eliminate on public benchmarks → confirm on your own 50–200-example golden set → hold the problem constant → let it settle. Leaderboard rank is a filter, not a verdict; the output is often a routing policy (which approach per task class), not a single winner. This skill decides; scouting hands off to research-arxiv-scout / research-scout, eval methodology to ai-evals. Full discover/vet/reframe process + emerging-classes (BLT, V-JEPA 2, Genie 3, diffusion-LM — "named but not yet load-bearing") in discovering-architectures.md.

Known Traps (Top Six)

The failure mode behind all of these: reaching for the familiar approach before classifying the problem. Full 19-trap + anti-pattern catalog in traps-and-anti-patterns.md.

• fine-tuning the wrong gap — fine-tune fixes behavior/format; missing knowledge → RAG, weak reasoning → a reasoning model / more thinking. Match the fix to the gap type.
• "trees always win on tabular" — false within a tabular-FM envelope (TabPFN-3 / TabICL v2); but quoting a stale size limit is its own trap — verify the current version.
• reflexive RAG / unchecked long-context — a static corpus may just fit in long context; but multi-fact long-context recall degrades silently (needle tests pass, real recall fails) — eval multi-fact, and price the cost cliff at volume.
• jumping to multi-agent when a single agent or linear tool-use workflow would do — multi-agent can lower quality on sequential tasks via coordination overhead.
• architecture-axis theatre — debating dense/MoE/SSM/diffusion for a builder who only calls a hosted API (it's invisible there); or treating embedding choice as throwaway config (it has re-index lock-in).
• silently dropping a candidate instead of eliminating it with a stated reason.

Core Principles

1. Classify before you choose. Output type + data type + constraints come first.
2. Start simple, promote on evidence. The default is the cheapest thing that can work; complexity must earn its place against a measured failure.
3. Eliminate with reasons, never silently. Every dropped option gets a one-line why.
4. Match the tool to the data. Tabular -> trees; sequence/language -> transformers; knowledge -> retrieval; actions -> agents.
5. Decide here, build there. This skill's output is a decision + handoff, not an implementation.

Navigation: Core References

• AI Skill Map — all 36 ai-* supporting deep skills grouped by decision lane; the handoff index
• Decision Matrices — cross-paradigm selection tables (classical ML vs DL vs LLM vs RAG vs fine-tune vs agent); also holds the transfer-knowledge, scale-axis, and emerging-classes tables
• When To Use What — per-approach "use it when / avoid it when" cheat sheet, including the GBDT family
• Discovering Architectures — discover → vet → reframe process for approaches newer than this skill catalogs
• Traps & Anti-Patterns — full 17-trap + anti-pattern catalog (the negative knowledge)

External Sources

See data/sources.json for primary references on model selection, the GBDT family, the prompt/RAG/fine-tune decision, and agent build-vs-not.

Fact-Checking

• Known bugs, regressions, framework/runtime footguns, and version-specific guidance must be verified against current primary web sources before being treated as current fact.
• Model rankings, library benchmarks, and pricing are volatile; verify current winners against official docs before recommending a specific one.
• If you cannot verify, present guidance as a dated assumption, not a fact.

Learnings Loop

Before applying this skill on a non-trivial task, read learnings.consolidated.md in this directory (and learnings.md if present).

After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.