The Autonomous
Research Scientist
OmniMind reads, understands, reasons over, verifies and extends heterogeneous research data. Not a chatbot over documents — a pipeline that decomposes a question, gathers evidence across every modality, argues from it, then attacks its own answer before showing it to you.
Built like a system, not a script
Clean architecture with ports and adapters. The domain layer imports no database and no vendor SDK; every driver is swappable by configuration.
Every Modality
PDFs, figures, charts, tables, equations, audio, video, websites and GitHub repositories are parsed into one retrievable index — assets become first-class chunks, so a question about a figure hits the figure.
Hybrid Retrieval
Dense and sparse retrievers run in parallel and fuse on rank, not score, because cosine and BM25 live on incomparable scales. Then listwise reranking, MMR diversification, and contextual compression.
Knowledge Graph
Eleven node types and ten edge types extracted with conservative entity resolution. PageRank and betweenness surface the bridges between subfields — where the unrun experiments live.
Verified, Not Asserted
Every claim is adjudicated in isolation against passages only. Hallucinated citation markers are stripped. Confidence is a computed function of verdicts, not a number the model chose.
Thirteen specialists, one blackboard
Agents never call each other. Each reads typed slots written upstream and writes its own, so the dependency graph is explicit, independent steps run concurrently, and any agent can be tested alone.
Planner
Decomposes the question into a typed agent DAG
Retriever
Hybrid dense + sparse with fusion and reranking
Vision
Reads figures, charts, diagrams, medical imagery
OCR
Recovers scans, screenshots and handwriting
Research
Pulls in external literature the corpus lacks
Knowledge Graph
Builds the typed entity graph
Reasoning
Synthesises evidence into an argued answer
Verification
Adjudicates each claim against its evidence
Citation
Binds claims to sources, audits coverage
Research Gap
Finds what the literature has not addressed
Experiment Design
Turns gaps into falsifiable protocols
Critic
Adversarial review, revises when warranted
Report
Assembles the long-form deliverable
Six stages, each independently measurable
Every stage can be switched off, which is what makes an ablation table possible rather than a claim.