Alberto Tuveri

SerSanInside SphereNodeLive at spherenode.com

Sphere AI

SphereNode's bilingual RAG mentor, grounded only in the proprietary curriculum and measured on a 161-question evaluation set.

A whole Sphere AI conversation in SphereNode, on dark green-black: a trader's question about managing risk after a drawdown; the answer, a general paragraph, then a paragraph headed What the transcripts actually support with numbered citation markers and a passage marked as a general concept the excerpts do not cover; the label Based on Spherenode sources; the cited lessons and the suggested lessons and live sessions as source cards; the AI disclaimer and the question box. An instructor's name and every lesson and session title are covered by flat grey bars.
Sphere AI, citing its sources A whole answer from the curriculum, with its sources, captured September 2026; a name and the lesson titles redacted.Visit

01 What it is

It asksbefore it guesses

Sphere AI is the bilingual (IT/EN) RAG mentor inside SphereNode. It answers only from the platform's proprietary curriculum: semantic chunking and embeddings in pgvector, tier-aware retrieval with a curated fast path for canonical answers, and a generation layer with an LLM relevance grader on top.

Stack

  • RAG
  • pgvector
  • Embeddings
  • Semantic chunking
  • LLM relevance grader
  • LLM evaluation
  • Guardrails
  • Supabase

02 The pipeline

Curriculumto answer

Every answer starts from the proprietary curriculum and passes a relevance grader before it reaches the trader.

  1. 01Proprietary curriculum
  2. 02Semantic chunking
  3. 03Embeddings in pgvector
  4. 04Tier-aware retrievalwith a curated fast path for canonical answers
  5. 05Generation
  6. 06LLM relevance grader

Then it

  • Answers, with citations
  • Asks for clarification
  • Refuses

Retrieval and generation From the curriculum to an answer, a question back, or a refusal.

03 Measured

Beforeand after

Every change was measured against the same bilingual evaluation set, with zero locked-content leaks and zero fabricated citations.

161-question bilingual evaluation set · before → after

  1. Answer-policy accuracy

    52%to68%

  2. Over-answering on questions to withholdLower is better

    56%to16%

  3. Clarification precision

    75%to94%

  4. Clarification recall

    32%to90%

  5. Unsupported claimsLower is better

    1.9%to0.6%

    Too small to draw on this scale.

  6. Citation correctness · zero fabricated citations

    100%

04 The guardrails

What itwill not do

The guardrails behind the numbers

  1. 01Asks for clarification instead of guessing.
  2. 02Refuses questions built on a false premise or asking for unsafe actions.
  3. 03Requires live evidence before answering anything time-sensitive.
  4. 04Never lets static FAQ content stand in for current market information.
  5. 05Reads the Trading Journal only as a consent-gated, owner-only source, personal data excluded by default.

05 The record

Studio
SerSanSersan LimitedLondonRemote
Role
Designed and built it, as Software & AI Engineer
Lives in
SphereNode, at spherenode.com
Evaluation
161 questions, Italian and English

More work

  1. 01SerSan Agent OSSerSan · multi-agent systemOpen
  2. 02SphereNodeSerSan · live at spherenode.comOpen
  3. 03Sphere AISerSan · RAG mentorThis case
  4. 04Maintech OMASerSan · workshop ERPNext
  5. 05sersan.ioSerSan · Jun – Sep 2026Open
  6. 06Domus TuaSerSan · real-estate brand platformOpen
  7. 07Brand filmSerSan · identity and filmOpen
  8. 08OblaPersonal project · multi-agent systemOpen
  9. 09Badante24hALS MCL Civitanova · freelanceOpen
  10. 10Voice AI AgentDOIT S.r.l · internshipOpen
  11. 11Supply ChainUniversity of CamerinoOpen