Skip to content
LearningNemo

NVIDIA-tool-first agentic platform

A CEO-led AI research swarm.

Learning Nemo is an agentic system where a reasoning CEO agent plans a task, decides whether to solve it directly or decompose it across specialist agents, checks its own work, and synthesizes a single governed deliverable — wrapped in memory, retrieval, evaluation, safety governance, and a real-time operator dashboard.

See the live dashboard How it works
35
modules
across 14 categories
14
functional categories
6-node
CEO decision loop
plan → gate
>85%
per-file test coverage
enforced in CI
~97
logged decisions
each traceable
~137
changelog increments

Design intentions

Three ideas shape everything

Learning Nemo is not a single product — it's a disciplined platform built around the NVIDIA ecosystem and designed to be re-pointed at new research domains.

Compounding improvement

NVIDIA-tool-first

Wherever a capability exists in the NVIDIA ecosystem, the swarm is designed around that tool rather than reimplementing it — NIMs and the NeMo toolchain for inference, retrieval, evaluation, curation, training, guardrails, and tracing. A deliberate benefit: as those tools improve, the swarm inherits the gains without re-architecting.

Platform, not a one-off

A reusable base, not a single product

The swarm is a modular platform meant to power multiple products across research-heavy domains. Its first application is Deep Medical Research Report Generation; the same core is designed to be re-pointed at additional research domains.

Swappable everywhere

Pluggable by construction

Orchestration, inference backends, vector storage, and generation backends are accessed through abstraction layers, so components swap without rewriting business logic. New orchestration strategies are added by registering a subclass — no changes to the CEO's core code.

How it works

A 6-node CEO decision loop

Most agentic systems are a single prompt-to-answer leap. The CEO instead runs a governed loop — and decides, per task, whether to work alone or spin up a team.

1

Plan

Decompose the task and decide serial vs. swarm.

2

Meta-critique

Stress-test the plan before acting.

3

Execute

Run directly or coordinate specialists.

4

Reflect

Loop back on low confidence.

5

Synthesize

Assemble a single governed deliverable.

6

Safety gate

Final governance + guardrail check.

Reflect loops back to plan whenever confidence is low — the CEO re-plans before it commits.
Walk through the orchestration

Capabilities

35 modules, 14 categories

Each capability is delivered by discrete modules with defined contracts and an honest status — including the parts still in integration testing.

Orchestration & Reasoning Core

3

The CEO plans, decides serial-vs-swarm, coordinates specialists, and gates its own output.

Inference & Precision

4

A dual inference path with per-role precision routing and multi-GPU/multi-host coordination.

Memory, Retrieval & Knowledge

5

A permanent, stratified knowledge base with production RAG and principle distillation.

Safety, Governance & Audit

4

Graduated autonomy, layered guardrails, and a tamper-evident audit trail.

Evaluation & Self-Improvement

7

Scoring, tournaments, training cascades, and prompt/parameter optimization.

In integration testing

Multimodal & Channels

4

Modality routing, parallel verification, generation backends, and multi-channel reach.

In integration testing

Tools, Prompts & Contracts

4

Tool discovery and creation, versioned prompts, and validated data contracts.

Observability, Operations & Foundation

4

The operator dashboard, tracing, configuration, and the CI/CD delivery pipeline.

First application · POC complete

Deep Medical Research reports

The proof of concept is complete: the swarm is generating full 40–70 page Deep Insight, hypothesis-generating medical reports today. The same intake → plan → gather → analyze → compose → gate → render pipeline is designed to extend to other domains.

  1. Step 1

    Intake & classification

    CEO (M004)

  2. Step 2

    Planning

    Research planner

  3. Step 3

    Evidence gathering ("BigBrain")

    Deep-research specialist

  4. Step 4

    Analysis

    Domain analyst panel

  5. Step 5

    Composition

    Report composer

  6. Step 6

    Quality gating

    Deterministic evaluator

  7. Step 7

    Charts & rendering

    Chart planner + renderers

  8. Full pipeline

The view inside the swarm

Watch the swarm think — read-only.

Most agentic systems are a black box. The operator dashboard is the opposite: live CEO reasoning, the orchestration flow, specialists as they spawn, the evidence debate, memory, governance, and system health — all on one surface. The public view is a point-in-time snapshot you can navigate.

Open swarm.learningnemo.com

How it's built

The discipline is the differentiator

Much of this system was built by AI coding agents under explicit guardrails. The program that produced it is its most distinctive feature.

Single source of truth

One canonical architecture document with enforced bidirectional contract integrity.

Arch → Spec → Code → Review → Merge

Every module follows the same gated lifecycle, with codified role-specific agent prompts.

Checking the agents' work

Explicit modes, minimal diffs, a living lessons-learned registry, and no silent fallback.

Everything documented

Numbered decisions, a versioned changelog, session logs, issues, RFCs, and root-cause analyses.

Gated, multi-tiered testing

>85% per-file coverage, plus deep E2E journeys validated on real GPU hardware.

CI/CD by design

GitHub Actions: test gate, profiling gate, contract consistency, deploy validation.

See the program-management methodology