Technology
NVIDIA-tool-first, abstraction-everywhere
The platform is built around NVIDIA NIMs and the NeMo toolchain. Every external dependency is governed by an allowlist and accessed through an abstraction layer where replaceability matters — so the platform is not locked to any single vendor component.
The stack
NVIDIA ecosystem
Leads the design- NVIDIA NIM (vLLM backend) — local inference
- NeMo Agent Toolkit — tracing / profiling
- NeMo Guardrails — safety validation
- NeMo Evaluator — scoring & judging
- NeMo Retriever — production RAG
- NeMo Curator — data curation
- NeMo Gym — self-play tournaments
- NeMo Framework (PEFT) — training
- TensorRT-LLM — precision controls
- RAPIDS (cuDF / cuML) — GPU data ops
- ASR / TTS NIMs — voice pipeline
Core platform
- Python 3.12 (pinned)
- Pydantic v2 — config & validation standard
- LangGraph — default orchestration backbone
- ChromaDB — vector store (1024-dim, behind abstraction)
Interface & operations
- NiceGUI — operator dashboard
- LiveKit — WebRTC voice pipeline
- Plotly — visualizations
- Prometheus — metrics
- Docker Compose + supervisord — deployment
- GitHub Actions — CI/CD
Compounding by design
The swarm gets better as the tools get better
Because Learning Nemo is designed around the NVIDIA ecosystem rather than reimplementing it, improvements in the underlying NIMs and NeMo toolchain flow into the platform without re-architecting. The design choice itself is a long-term capability strategy.