Solo builder · AI Engineer · 2025
Sophia — AI Assistant
A conversational AI research playground for memory, tool-use and multi-step planning.
PythonLLMsPrompt EngineeringAgents

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
Prompt orchestration
Structured system prompts and templates separate role, tools and task state so behaviour stays predictable.
- 2
Tool calling
The model returns structured calls into a small tool set instead of free-form instructions.
- 3
Lightweight memory
Context is stored as structured JSON state, which is readable and diffable rather than opaque.
- 4
Planning loop
Multi-step tasks are decomposed, executed and re-checked against the original request.

Tech stack
Python
LLMs
Prompt Engineering
Agents

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