Public: Intercom Fin customer-support agent

Architectural notes

Fin's answering model is no longer a catalog model and is omitted: Fin 1
(2023) ran on OpenAI GPT, Fin 2 (Oct 2024) was "powered by Anthropic's Claude"
(Sonnet 4.0 per Intercom), and the current Fin Apex (Mar 2026) is an in-house
model post-trained on proprietary support data atop an undisclosed open-weights
base - none map cleanly to the catalog.

Per their fin.ai research posts, retrieval is a custom in-house RAG: a fine-tuned
retrieval model on a Snowflake Arctic Embed 2 base plus a ModernBERT reranker
(fin-cx-reranker). Voyage-3-large and Cohere Rerank v3.5 appear only as benchmark
baselines, not in production; the vector store is not publicly named.

Intercom ships a first-party MCP server (Cloudflare-hosted) and Fin also acts as
an MCP client. Fin Voice uses an STT-to-LM-to-TTS pipeline (OpenAI Realtime API
for prototyping; STT/TTS vendors undisclosed). Evaluation is via large-scale A/B
tests, LLM-as-judge, and an internal FinRank benchmark. No third-party agent
framework, eval, or observability tool is publicly named.