Combining autonomous research agents, mathematical search, and formal verification into a single research workspace to solve the unsolved!
Built at Hack the North 2026 on 4h of sleep:
- Best Use of Vultr
- 2nd Best Use of Devin
Checkout the Devpost submission: click
-
Multi-Agent Mathematical Research
- openJiuwen WorkSwarm supports separate research agent branches in parallel during exploration
- Coordinates 3 researchers, 1 challenger, 1 coordinator, and 1 proof-writer agent
- Supports fresh research directions after getting stuck or realizing failure
-
Formal Proof Verification
- Candidate proofs are checked using Lean rather than accepted directly from challenger outputs
- Lean feedback is returned to the research workflow for further refinement
- Challenger experiment worker runs bounded exact-arithmetic searches on Vultr infrastructure to reduce forced LLM computations
- Lean checker also runs on a Vultr-hosted VM to remove user download requirements
-
Pre-research Knowledge Base
- Discovers mathematical literature through OpenAlex across 16 research areas
- Ranks papers by citation count and stores metadata
- Deduplicates papers and preserves source links including arXiv/PDF references
- Generates embeddings for semantic/dense retrieval with vector search
-
Persistent Research Infrastructure
- MongoDB stores research state, mathematical knowledge, graph relationships, and artifacts
- Redis manages research projects and ingestion jobs
- Research API creates and tracks research episodes
- Background workers execute literature retrieval, small arithmetic experiments, and verification
- TypeScript
- React / Next.js
- Node.js
- Python
- openJiuwen SwarmFlow
- MongoDB + MongoDB Atlas Vector Search
- Redis
- Lean 4
- OpenAlex
- Vultr
- Docker
The full local setup requires an unusual amount of API keys, external services, and remote compute infrastructure, so a setup instruction won't be provided...
But check out the complete platform through our hosted deployment! Try Triviality →
Triviality/
├── apps/
│ ├── web/
│ ├── research-api/
│ ├── research-worker/ # Research orchestration + persistence
│ │ ├── src/
│ │ └── tests/
│ ├── research-swarm/ # WorkSwarm / SwarmFlow runtime
│ ├── experiment-worker/ # Remote computational experiments
│ ├── paper-ingest/
│ ├── api-papers/
│ ├── api-auth/
│ └── api-billing/
│
├── packages/
│ ├── database/ # MongoDB schemas + persistence
│ ├── services/
│ ├── ui/
│ └── typescript-config/
│
├── swarm-skills/
│ └── math-research/
│ ├── roles/ # Research agent definitions
│ ├── scripts/ # Workflow + Lean checking
│ ├── SKILL.md
│ └── workflow.md
│
├── backend/
│ ├── artifacts/
│ └── benchmarks/
│
├── deploy/
│ └── vm/ # Vultr VM deployment
│
├── config/
│ └── research-models.json
│
├── docs/
├── scripts/
├── external/
│
├── docker-compose.yml # MongoDB + Redis
├── pnpm-workspace.yaml
├── turbo.json
└── package.json
Supported by integrating Huawei openJiuwen's WorkSwarm, our research workflow looks like:
Researchers
▼
Challenger / Iteration
▼
Reviewed Complete Argument
▼
Proof Writer
▼
Mathematical Proof + Lean Formalization
▼
Lean Checker
╱ ╲
Fail Pass
▼ ▼
Research Verified Proof
Loop ▼
Research Worker
▼
proofDocument(...)
▼
LaTeX Output
Research episodes preserve intermediate attempts into an episode-sepcific knowledge pool and failed verifications are returned to the workflow so agents can revise or abandon the candidate.
- Designing multi-agent workflows for open-ended mathematical research
- Connecting LLM-generated reasoning with deterministic computational evidence
- Integrating Lean formal verification into an autonomous research loop
- Running remote experiment and verification workloads on cloud compute