Evidence-backed AI stack directory
The AI stacks of real companies, in production — with receipts.
Which models, tools, and infra real companies actually run — not another AI tools list. Every fact is backed by a source link, a verbatim quote, and a last-verified date.
- companies
- 100
- verified facts
- 222
- receipts
- 284
- stack layers
- 10
Company stacks
Grouped by layer, receipts on every entry.
1up
1 facts · 1 receipts
AI answer engine automating RFPs and security questionnaires for sales teams.
Verified Jul 27, 2026
Allspice
1 facts · 1 receipts
AI cooking tools for recipe publishers.
Verified Jul 27, 2026
Amgen
1 facts · 1 receipts
Global biotechnology company developing therapeutics.
Verified Jul 27, 2026
Applied Compute
1 facts · 1 receipts
Startup building reinforcement-learning-trained enterprise AI models.
Verified Jul 27, 2026
Aquant
1 facts · 1 receipts
Agentic AI platform for servicing complex equipment.
Verified Jul 27, 2026
Avalara
1 facts · 1 receipts
Tax compliance automation software company.
Verified Jul 27, 2026
Ten layers of the production AI stack
Every fact maps to exactly one layer of the taxonomy.
Models & Providers
31 factsFoundation model providers and specific models used (OpenAI, Anthropic, Gemini, Llama, Mistral, self-trained).
Inference & Serving
36 factsWhere models run — clouds (Bedrock, Vertex), GPU providers, inference platforms (Together, Fireworks, Baseten, Modal), self-hosted serving (vLLM).
Orchestration & Agents
23 factsAgent frameworks, LLM app frameworks, workflow engines (LangChain/LangGraph, LlamaIndex, Temporal-for-AI, MCP usage, custom).
RAG & Vector Search
32 factsVector databases, embeddings pipelines, retrieval infra (Pinecone, Weaviate, Qdrant, pgvector, Turbopuffer, Elasticsearch-for-RAG).
Evals & Observability
28 factsEval frameworks, LLM tracing/monitoring (Braintrust, LangSmith, Langfuse, Arize, W&B Weave, custom evals).
Fine-tuning & Training
16 factsTraining/fine-tuning stack — frameworks, compute, RLHF tooling, model post-training.
Data & Labeling
4 factsLabeling vendors and data pipelines for AI (Scale, Surge, Labelbox, synthetic data tooling).
Guardrails & AI Security
12 factsSafety filters, prompt-injection defense, PII redaction, AI security vendors and practices.
AI Coding Tools
24 factsWhat their engineers code with (Cursor, Claude Code, Copilot, Windsurf, internal assistants).
Internal AI Productivity
16 factsCompany-wide AI tools for non-engineering work (ChatGPT Enterprise, Glean, Notion AI, meeting AI, support AI).
A fact without evidence does not ship
AI stacks are backend-invisible — no crawler can detect them. So every entry here is built from what companies say in public, held to a strict standard.
Primary sources only
Job postings, subprocessor lists, engineering blogs, vendor case studies, conference talks, public repos — never frontend guesswork.
Verbatim receipts
Every fact carries the source URL, its publication date, and a word-for-word quote that supports the claim. No paraphrasing, ever.
Graded, then gated
Facts are graded high or medium confidence. Anything weaker is dropped, not published. Tools a company moved off stay visible, marked historical.
Re-verified on schedule
Every source link is re-checked and each fact shows its last-verified date. Stale directories die; this one is built to stay fresh.
Full sourcing rules on the methodology page.
Run in the open, by an agent
This directory is ~90% AI-operated: watchers find candidate sources, a verification gate tries to refute every fact before it publishes, and a human reviews the queue. The ops log is public.