Solo builder · AI Engineer · 2025

Sophia — AI Assistant

A conversational AI research playground for memory, tool-use and multi-step planning.

PythonLLMsPrompt EngineeringAgents
Sophia — AI Assistant — A conversational AI research playground for memory, tool-use and multi-step planning.

Goals

  • Test whether an LLM can hold multi-turn task state reliably.
  • Give the assistant tools rather than only text output.
  • Keep memory inspectable instead of hidden in an embedding blob.
  • Build the agentic patterns later reused in J.A.R.V.I.S — NIX.

Approach & architecture

  1. 1

    Prompt orchestration

    Structured system prompts and templates separate role, tools and task state so behaviour stays predictable.

  2. 2

    Tool calling

    The model returns structured calls into a small tool set instead of free-form instructions.

  3. 3

    Lightweight memory

    Context is stored as structured JSON state, which is readable and diffable rather than opaque.

  4. 4

    Planning loop

    Multi-step tasks are decomposed, executed and re-checked against the original request.

Agent loop diagram: user request, structured system prompt, LLM reasoning engine, tool call registry, JSON memory state, planning step and a re-check and validate stage looping back until the goal is met.
Agent loop — structured prompting, tool calls, JSON memory and a re-check stage.

Tech stack

Python
LLMs
Prompt Engineering
Agents
Developer tool interface with a multi-turn agent chat transcript including tool calls on the left and the inspectable JSON memory state with goal, facts, plan and status keys on the right.
Inspectable memory — the agent's JSON state alongside the conversation.

Results

Multi-turn

task flow with recall and citations

JSON

inspectable memory state instead of opaque embeddings

Reused

patterns carried into J.A.R.V.I.S — NIX

Frequently asked questions

More case studies

All projects