Guglielmo Paesano AI engineer and founder of AIGROW

I build AI systems that do people's work, and I measure where they still need people.

contact@guglielmo.ioDiscord, 4K members

SocrateOS

The production system of 1o1.org, a dialectic AI platform for reflective decision making.

SocrateOS is the production system of 1o1.org, a nonprofit building a dialectic AI platform for reflective decision making. It helps a person reason through a decision by questioning it, at three depths, and it remembers what was established between sessions.

Status
In production
Role
Technical partner and lead AI engineer
Year
2026
Live at
1o1.org
SocrateOS: 1o1.org, filmed as it runs

What it is

The production system I designed, built and run for 1o1.org, a nonprofit founded by Romanus Berg: a dialectic AI that helps people reason through a decision by questioning it, in three depths (Ping, Sprint and Chat), with memory across sessions, retrieval over the user's own documents and safety checks before any model call. It is the deployment behind my paper at the HAIC workshop, NeurIPS 2026.

Visit 1o1.org

Built with

  • Next.js 16
  • FastAPI
  • PostgreSQL 18
  • pgvector
  • Stripe
  • Docker Swarm

What you can check

  • Picked from 44 proposals; a working MVP in ten days
  • About 370 automated tests, CI on every push, isolated staging
  • Retrieval over each user's own documents on pgvector with HNSW

What a dialectic AI is

A dialectic AI helps someone reason through a decision by questioning it. An assistant completes the request it is given. This system asks what the request assumes, presses on the weakest step, and leaves the person with a position they have tested. 1o1.org exists for reflective decision making, and that purpose shapes every part of the build.

Most AI products are tuned for agreement, which makes them pleasant and rarely changes anyone's mind. A system built to question has to hold a model of what the person believes across turns, notice when a new statement contradicts an earlier one, and say so.

The memory, the retrieval and the safety checks described below all follow from that requirement.

Three depths: Ping, Sprint and Chat

The platform meets a person at three depths. Ping is one public exchange with no sign in. Sprint is a five step dialectic that ends in a saved artifact. Chat is open ended and keeps its memory across sessions. Each depth asks for a different amount of time, and a person can stop at any of them.

Ping lets someone meet the method before creating an account. Sprint gives the dialectic a fixed shape: five steps, then something to keep. Chat is for the decisions that take weeks, where the value is that the conversation continues from what was already established.

The same engine runs under all three: multi step dialectic flows, with persona pressure so that a question meets argument from more than one direction.

Memory and retrieval on one database

Memory is a persistent graph of what the person has established, and retrieval runs over the person's own documents with pgvector and an HNSW index. Both live in PostgreSQL 18 next to the relational data, so there is one database to back up, one permission model, and no job keeping a vector store in step with the system of record.

A transcript search returns the moment something was said. A graph returns what a statement implies and what it contradicts, which is what a conversation about a decision needs.

Keeping the vectors in Postgres was a deliberate trade. A dedicated vector database is faster at very large scale. At this scale, running two stores and keeping them consistent costs more than it buys.

RAG on PostgreSQL with pgvector, as a service

Safety before any model call

Harmful input is stopped before any model is called. Every claim the AI makes is labelled as fact, inference or speculation, and when a model provider fails, the person is shown the failure as it is. Someone reasoning about a real decision should always know what kind of statement they are reading.

Stopping harmful input before the model call means the check does not depend on the model behaving well. It is ordinary code with tests, and it runs first.

The labels matter for the same reason. A person weighing a decision will treat a confident sentence as true unless told otherwise, so the system says which sentences are established, which are reasoned and which are guesses.

From ten days to production

1o1.org picked the proposal from 44. A working MVP of the dialogue engine existed ten days later, and the product was then built milestone by milestone: Next.js 16 and FastAPI on PostgreSQL 18, Stripe, Docker Swarm and Traefik behind Cloudflare, about 370 automated tests, CI on every push and an isolated staging environment.

Each milestone ended in something the client could run and check before the next began. That is how a system of this size stays correctable: a misunderstanding costs one milestone.

The founder, Romanus Berg, wrote in a public recommendation that the architecture is clean enough that a student cloned it and ran it independently. The full text is on the home page, with its source.

How an engagement runs

The research that came out of it

The deployment is the subject of a paper, "Instrument before you evaluate: write-path invariants for deployed human-agent systems", written with Romanus Berg and accepted at the Human-AI Coevolution (HAIC) workshop at NeurIPS 2026. It argues that whether a human-agent system can be evaluated is decided by its write path: what it records, with what attribution, and what it may overwrite.

The paper is a field report. It comes from running this system with real users, and it closes with a checklist a team can answer before it designs an evaluation.

What the paper says

Questions

What makes this different from a chatbot with long-term memory?
A chatbot with memory stores past messages. Here the memory is a graph of what the person has established and how it connects, and the method is dialectic: the system questions a position where an assistant would complete the request.
Why pgvector rather than a dedicated vector database?
Because the memory graph, the documents and the embeddings belong to the same data, and splitting them across two systems buys a synchronisation problem. PostgreSQL with pgvector and an HNSW index keeps retrieval and relational structure in one place, which matters more than raw vector throughput at this scale.
How is harmful input handled?
A check runs before any model call and stops it. The check is ordinary code with its own tests, so it does not rely on a model refusing. Every claim the AI makes afterwards is labelled as fact, inference or speculation.
How was the work structured with the client?
As approved milestones, each ending in something the client could run and verify before the next began. The first working version existed in ten days, and the engagement has continued since April 2026.

Facts on this page verified 2026-11-10

Tell me what you need built.

Write what the system has to do, what exists already and when it has to be live. I read every brief myself and answer in writing, with a scope and a price for each milestone.

Or write to contact@guglielmo.io

The human check appears here.

No calls and no sales sequence. One written answer from me.