Inspiration
Research rarely happens in a straight line. Observations, half-formed hypotheses, failed attempts, and open questions pile up in notebooks, and the structure that connects them lives mostly in the researcher's head. We wanted a system where a researcher could keep writing freely while something alongside them reconstructed, challenged, and tested the understanding hidden in those notes — without ever quietly rewriting the scientific record. Our guiding line: the researcher writes the history. The system helps reconstruct, challenge, test, and evolve the understanding.
What it does
VectorHeart turns unstructured research notes into an evolving, inspectable body of scientific understanding.
- Research Journal: free-form, dated entries for observations, questions, hypotheses, experiment notes, and uncertainty, with autosave, revision history, tags, and stale-edit protection.
- Research Graph: a visual map of hypotheses, observations, questions, experiments, results, explanations, claims, and uncertainties, linked by typed relationships (supports, contradicts, explains, tests, motivates, depends on). Graph generation produces a preview for review. Nothing is written to the record until the researcher accepts it, and every correction, rejected link, and rationale is preserved.
- AI Accelerate: asks a model to construct hypotheses, connect ideas, challenge assumptions, modify formulations, and propose discriminating experiments. AI contributions land directly on the graph as provisional, individually reviewable nodes and edges, visibly distinct from researcher-authored records.
- Bounded Autonomous Investigation: a persisted agent that reasons through up to five steps, selects only from registered scientific tools (never arbitrary code), pauses for explicit human approval before executing, runs operations through the Scientific Kernel, and stores real results for follow-up reasoning.
Throughout, VectorHeart keeps five things separate: what the researcher recorded, what the model inferred, what the AI proposed, what external papers reported, and what the system actually measured.
How we built it
- Frontend: a lightweight web app with a project workspace, Notebook, Canvas/Research Graph (pan/zoom, node selection, saved positions), and Research Intelligence surfaces.
- Backend: Python with persistent SQLite and Alembic-managed schema migrations. Everything is persisted: research objects, relationships, revisions, conversations, proposals, journal entries, graph corrections, and investigation history.
- Model layer: provider adapters for OpenAI, Google Gemini, and Anthropic using each SDK's structured-output interface. Provider and model are set per project, and every proposal carries provider/model and context-revision provenance.
- Scientific execution: registered tool interfaces, draft operations, approval, execution, evidence recording, and resumable investigation state with checkpoints.
- Safeguards: project isolation, validation, stale-revision checks, duplicate-request protection, and hard approval boundaries between proposal acceptance and experiment execution.
Challenges we ran into
-Building something beyond an AI chatbot. We wanted scientific understanding itself to become computationally accessible: a living network of hypotheses, experiments, evidence, contradictions, and revisions. Transforming free-form research journals into editable knowledge graphs while preserving the original scientific record required substantial architectural work.
=Balancing autonomy with scientific integrity. AI-generated hypotheses aren't verified discoveries. We implemented reviewable proposals, human approval, versioned scientific objects, and persistent corrections. Enabling autonomous investigations to pause for approval and resume across processes without duplicating experiments was particularly challenging.
=Limited access to powerful AI inference. Without paid API capacity or dedicated infrastructure, we relied on accessible Gemini models. Repeated HTTP 503 overload errors interrupted testing, while lighter models constrained the sophistication of our demonstrations. We designed VECTORHEART to support interchangeable AI providers rather than depend on any single model.
Accomplishments that we're proud of
We built the foundations of a scientific intelligence system, not another AI research assistant.
-VECTORHEART transforms research journals into living knowledge graphs containing hypotheses, observations, questions, experiments, and scientific relationships. Researchers can correct AI interpretations, modify connections, and preserve how their understanding evolves.
-Our AI Accelerate system proposes new hypotheses, alternative explanations, relationships, and experiments directly within the graph. Every suggestion has an inspectable rationale and remains subject to researcher approval.
-We also developed a persistent Scientific Kernel supporting real numerical experiments, evidence tracking, provenance, and bounded autonomous investigations with human approval. Literature Intelligence connects research questions to actual scholarly publications and scientific evidence.
-Our model-independent architecture integrates Google Gemini, OpenAI, and Anthropic. At our Stage 3 checkpoint, we achieved 63 passing backend tests, 21 frontend tests, successful live Gemini graph generation, and verified cross-process experimental execution using NumPy.
-Our proudest accomplishment is turning research from a collection of static documents into an evolving, inspectable, experimentally challengeable system of scientific understanding.
What we learned
AI's greatest contribution to science may not be answering questions, but helping researchers discover which questions deserve asking. Scientific journals contain observations, failed experiments, changing assumptions, and intuitions that rarely survive into publications. Preserving this intellectual history could fundamentally change how research progresses.
We also learned that scientific integrity must be architectural. Separating speculation from evidence, AI interpretations from researcher observations, and approval from experiment execution is essential for trustworthy autonomous research.
Our biggest realization concerns the future of university research infrastructure.
Imagine universities operating their own GPU clusters running powerful open-weight scientific AI models, connected to laboratories, research databases, and systems like VECTORHEART. Researchers could access sophisticated AI capabilities while keeping sensitive scientific data within institutionally controlled infrastructure, subject to appropriate security and access controls.
Rather than depending entirely on expensive commercial APIs, universities could build shared, privacy-conscious scientific intelligence infrastructure that grows alongside their research.
VECTORHEART's model-independent architecture is an early step toward that possibility.
What's next for VectorHeart
1. Research simulations for students. A researcher's or professor's work is often opaque to anyone who wasn't there for it. Because VectorHeart already holds the journal, the graph, and the actual results of a project, we can turn that record into an interactive simulation: a student steps through the hypotheses, sees why each experiment was run, is asked what they would try next, and compares their choice to what actually happened. The goal is to make a research project something you can learn by walking through and then contribute to, with the student's proposals landing on the same graph as provisional, reviewable nodes.
2. Thesis writing from the graph. A thesis is largely a reconstruction of a research history that already exists in scattered notes. With every observation, hypothesis, rejected link, experiment, and result stored with provenance, VectorHeart can assemble a first draft directly from the record (methods from executed operations, results from stored evidence, discussion from the accepted and rejected proposals) so the researcher spends their time arguing, not excavating.
3. Collaboration. Multiple researchers working on one shared graph, with per-author attribution on every node and edge, reviewable proposals from collaborators handled the same way AI proposals are today, and a clear record of who accepted or rejected what and why.
4. Time travel.: Time Travel can be a future feature: VECTORHEART's ability to reconstruct, navigate, and compare historical states of a researcher's scientific knowledge, preserving the evolution of hypotheses, evidence, experiments, decisions, and interpretations over time.
Built With
- agentic
- alembic
- gemini
- javascript
- knowledge-graph
- llm
- math
- numpy
- pytest
- python
- sqlite
