🚀 Proof-of-concept for Step S1 — Secure Usage Data Collection — from the NVIDIA paper:
Small Language Models are the Future of Agentic AI
Modern multi-agent systems increasingly depend on continuous instrumentation — not just for observability, but to support downstream adaptation. Usage logs are the foundation for training specialized Small Language Models (SLMs) that power agent capabilities like reasoning, extraction, and summarization.
This project focuses on Step S1 from the NVIDIA pipeline: collecting and securing high-quality logs of agent interactions, tool usage, and outputs.
🔐 Logging is essential to enable fine-tuning SLMs via task clustering, selection, and specialization (Steps S2–S5).
📌 Architecture reference (via Mistral.ai):
Based on the NVIDIA paper, Step S1 includes:
- Logging of:
- Agent prompts (non-HCI only)
- Output completions/responses
- Tool call inputs and outputs
- Optional latency or timing metadata
- Storage in a secure, encrypted pipeline
- Anonymization of user or org-specific content (PII, PHI)
| Layer | Tools/Options |
|---|---|
| Language | Python 3.11+ |
| Logging | structlog or loguru |
| Anonymization | spaCy, Presidio, or rule-based masking |
| Storage | JSONL, SQLite (encrypted), or S3 later |
| App Layer | FastAPI or LLM framework (for future multi-agent demo) |
- Define core logging decorator
- Add anonymization middleware (rule-based or NLP)
- Secure encrypted local storage backend
- Instrument toy multi-agent app (e.g., chat + tool use)
- Collect dataset for task clustering (→ Step S3)
- NVIDIA (2024): Small Language Models are the Future of Agentic AI
- Mistral.ai: Agents API Blog Post