Built on LangGraph — a compiled StateGraph with persistent SQLite checkpointing. The agent loops, pausing at every tool call for human approval before execution.
Router
Classifies request — LLM path or direct simulation
Clarification
Asks pre-run questions if the task is ambiguous
Task Planner
Breaks request into a structured numbered execution plan
LLM Agent
Reasons over the plan, selects tools, generates code
CORE
Human Approval
Graph pauses — you inspect the tool call and confirm
HITL
Tool Executor
Runs the approved tool call, streams result back
EXEC
↑loops back to LLM Agent — repeats until task complete
Persistent Checkpointing
Every graph step is saved to SQLite (dev) or PostgreSQL (prod). Pause mid-task, restart the server, resume exactly where you left off.
LangGraphSQLite → Postgrespause · resume
WebSocket Streaming
Agent logs, tool outputs, and artifacts stream to the UI in real time. Every node transition is visible as it happens.
FastAPI WebSocketreal-time logs
Dual-Scope RAG
Session-scoped index over uploaded files + project knowledge base. Semantic search with ChromaDB, embeddings via OpenAI or a local model.
ChromaDBOpenAI / localsession + project
Persistent Jupyter Kernel
A long-lived kernel per session — variables, imports, and DataFrames survive across tool calls. Full numpy / scipy / matplotlib stack.
Jupyternumpy · scipymatplotlib
Tools exposed to the LLM Agent
run_code
Execute Python in the persistent Jupyter kernel. Read and write session files.
query_knowledge
Semantic RAG retrieval from session files and the project knowledge base.
isodistort
Crystallographic analysis — ISODISTORT, PSEUDO, AMPLIMODES, SYMMODES, SUBGROUPS via Bilbao server.
convert_cif
CIF crystal structure conversion, symmetry transformations, and format normalisation.
Capabilities
Everything your lab needs
A complete AI research stack — from reading instruments to submitting manuscripts.
Data Ingestion
15+ scientific formats parsed automatically: instrument logs, CIF, HDF5, MATLAB, NI TDMS, NetCDF, OpenFOAM — no configuration.
CSV/TSVHDF5CIFMATLABNI TDMSNetCDFOpenFOAM
Physics-Aware Validation
RAG knowledge base + ReAct agents verify physical plausibility. The system understands domain physics, not just numbers.
ArXiv RAGPhysical rangesUnit checksSimulations
Automated LaTeX Reports
Full report generation: sections, figures with captions, tables from DataFrames, equations, BibTeX bibliography — compiled to PDF.
pdflatexBibTeXTikZEquations
Human-in-the-Loop
The graph pauses before every tool execution. Inspect the generated code, edit it, approve or cancel — then resume from any checkpoint.
Code editorStep rollbackCheckpoint resume
Scientific Database Search
Semantic search across arXiv, Materials Project, ICSD, and your own uploaded knowledge base. Results cited inline in the final report.
arXivMaterials ProjectICSDCustom RAG
ML Analysis & Fitting
Curve fitting, PCA, spectral decomposition, FFT, regression, anomaly detection — all executed in the persistent Jupyter kernel.
Curve fittingPCAFFT / PSDRegression
Pricing
Start free, scale when ready
No credit card required. Upgrade once LabTeX AI has saved you your first hours.