Skip to content

N-Way Self-Deliberating Evaluation Engine core for AI agent orchestration

License

Notifications You must be signed in to change notification settings

peeramid-labs/nsed

Repository files navigation

NSED

CI Coverage License: BSL 1.1

Enterprise-ready agentic orchestration layer implementing the N-Way Self-Evaluating Deliberation (NSED) protocol. Multiple AI agents work on your task in parallel, having intermediate checkpoints and converging on a final answer. NSED delivers frontier reasoning ability through a Mixture-of-Models architecture that amplifies the strengths of individual models. It shows less sycophancy, cost tracking, optimisation, and a full compliance audit trail managed through NATS as a high-performance communication bus.

Why NSED

Three small open-weight models (20B, 8B, 12B) on consumer hardware score 84% on AIME 2025 through NSED deliberation — matching DeepSeek-R1 and coming within 1 point of GPT-5. The same models score 54% with naive majority voting. (paper)

Problem How NSED Solves It
Single-model quality ceiling Mixture of Models: agents from different providers cross-evaluate each other every round — errors get caught, reasoning improves iteratively
Black-box AI decisions Full audit trail: every proposal, evaluation, score, and reasoning trace is persisted and streamable via SSE
Data leaves your network Run entirely on your infrastructure with local models (Ollama, vLLM) — documents never touch a third-party API
No human oversight at runtime Inject guidance mid-deliberation, expose custom tools agents can call back to your systems, set time/token/cost budgets
Vendor lock-in Provider-agnostic: mix Ollama, vLLM, Together AI, or any compatible endpoint in the same session
Frontier model hardware costs A swarm of 8B–20B models in 64 GB total VRAM matches reasoning quality that normally requires 80–140 GB for a single monolithic model

Quick Start

git clone https://github.com/peeramid-labs/nsed.git && cd nsed
cp .env.example .env          # add your API keys (Together AI, etc.)
make setup                    # install tools, configure git hooks
make docker-up                # start NATS + Ollama infrastructure
make run                      # build and run the orchestrator
INFO  Dashboard  → http://127.0.0.1:8080/dashboard
INFO  Swagger UI → http://127.0.0.1:8080/swagger-ui/

Or try simulation mode (no API keys needed): make dev-sim

For binary installation, Docker, Homebrew, and APT options see the orchestrator README.

Workspace

nsed-agent-sdk  <--  nsed-agent  <--  nsed-orchestrator
    traits           implementations        server
Crate Description
nsed-agent-sdk Trait definitions and data types for building custom agents
nsed-agent Reference agent implementation with NATS worker runtime
nsed-orchestrator API gateway, job broker, deliberation engine, dashboard
nsed-benchmarks Benchmark suite (GSM8K, AIME, LiveCodeBench)

Each crate has its own Makefile with crate-specific targets (e.g. make run-agent, make test-sdk). The root Makefile includes all four, so every target works from the workspace root. See docs/development.md for the full list.

Documentation

Topic Link
Architecture docs/architecture.md
Agent Development docs/agent-development.md
Agent Internals (Library API) docs/agent-internals.md
Authentication docs/auth.md
NATS Integration docs/NATS.md
Development & Configuration docs/development.md
Contributing CONTRIBUTING.md
API (Swagger UI) http://localhost:8080/swagger-ui/ (while running)

Reproduce paper results

To reproduce results we've obtained during paper research please switch on to paper-reproduce branch of this repository which is frozen on the old commit "as is". Further guidelines are in the readme of that branch.

License

Component License
crates/nsed-agent-sdk/ MIT
crates/nsed-agent/ BSL 1.1
crates/nsed-orchestrator/ BSL 1.1
crates/nsed-benchmarks/ Apache 2.0

Source-available under the Business Source License 1.1. Free for entities under $1M USD annual revenue, research, education, and development. Commercial license required above the threshold. Converts to AGPL-3.0 after 4 years.

See NOTICE for patent, AI training, and ethical use terms.


Copyright (c) 2025 Tri Ratna Technologies Limited, trading as Peeramid Labs.

About

N-Way Self-Deliberating Evaluation Engine core for AI agent orchestration

Topics

Resources

License

Contributing

Stars

Watchers

Forks

Packages

No packages published

Contributors 2

  •  
  •