AI/ML · Data Engineering · Full Stack
NEXUS AI
Multi-agent research and execution workspace
- Status
- Complete · actively extended
- Year
- 2026
- Live application
- Runs locally with Docker Compose and Ollama. Not publicly deployed.

- citation grounding checks
- 6/6
- citation grounding checks
- unsafe SQL statements blocked
- 8/8
- unsafe SQL statements blocked
- canary leaks
- 0/5
- canary leaks
01 Problem
Most AI apps return one opaque answer. NEXUS persists everything that produced it: the plan, the agents that ran, every tool call, the cited passage, the query behind each figure and a verification report - so a result can be inspected, replayed and exported.
02 Product overview
Missions run as a LangGraph graph: a planner proposes steps, a human approves, research, document and data agents execute, then a synthesizer, an evidence verifier and an artifact generator finish the job. Missions can pause, resume, cancel and retry from a checkpoint. The workspace also includes a knowledge library with hybrid retrieval, a DuckDB data lab, an artifact studio, a workflow builder and an authenticated MCP server.
Implemented
- Streaming chat workspace with four modes
- Mission Control: Planner → Human approval → Research / Document / Data agents → Synthesizer → Evidence verifier → Artifact generator
- Pause, resume, cancel and retry-from-checkpoint via a Postgres checkpointer
- Knowledge library with pgvector + full-text retrieval fused by RRF, plus a PDF viewer
- Data Lab with a DuckDB sandbox, guarded SQL and typed operations
- Authenticated MCP server over Streamable HTTP
Planned, not yet built
- Consuming external MCP servers
- OCR
- NLI-based verification
- Reranker
- Multi-worker scaling
- OpenTelemetry
Credit: Evaluation quality gates were contributed by collaborator rithika-sr (merged PR #1).
03 System architecture
04 Core technologies
Agents & AI
- LangGraph
- Ollama (qwen3:4b-instruct)
- nomic-embed-text
- MCP
- SearXNG
Data
- PostgreSQL + pgvector
- DuckDB
- sqlglot
- Redis
- arq
Application
- FastAPI
- Next.js 16
- React 19
- Monaco
- Radix
05 Engineering challenges
Small models, structured output
A 4B local model doesn't reliably emit valid plans. Constrained decoding plus a single repair retry made outputs dependable without a bigger model.
Plans that stay on the rails
Instead of free-form plans, the analyst agent selects from a derived catalog of operations, which keeps execution safe and testable.
Durable human approval
Missions pause for approval and must survive restarts. A LangGraph Postgres checkpointer makes interrupts durable.
06 Technical decisions
- 01Deterministic evidence verifier instead of an LLM judge
- 02Redis streams for chat; Postgres plus pub/sub for tasks
- 03Self-hosted SearXNG for web research
- 04No arbitrary code execution anywhere in the agent loop
07 Verified test results
- Backend
- 146
- Frontend (Vitest)
- 20
- Offline
- 49
- Playwright journey
- 1
Latest GitHub Actions run on main: passing. Counts as reported in the repository's README and testing docs.
08 Screenshots
09 Links
No public deployment. Runs locally with Docker Compose and Ollama. Not publicly deployed.


