Inspiration
Mechanochemistry can change reaction outcomes with shear and confinement, but most “AI scientist” demos are too open-ended to trust. We wanted a narrow, defensible closed loop: read a small literature corpus, propose one bounded testable question, run only allowed ASE/UMA simulations, and write findings back into memory — without claiming global novelty.
What it does
Mechanist is a bounded research agent for Diels–Alder mechanochemistry:
Ingest literature into Cognee and extract structured evidence Propose a gap question (deterministic MVP; Bedrock/Strands-ready) and optionally check the live web with Bright Data Create + validate experiments against simulation_rules.yaml Stage MD jobs for ASE/UMA-OMol (local stub today; HPC via SSH/Slurm designed) Analyze trajectories (favorable configurations, distances, orientation) and return honest scientific feedback Show the loop in a Streamlit demo (knowledge → gap → trajectory GIF/XYZ → analysis)
How we built it
Python library (mechanist/) for memory, literature, validation, staging, MD runner, analysis Cognee for long-term research memory (add → cognify → search) Bright Data MCP for current public literature checks Hard safety rail: parameter validation before any run ASE deformation MD with injectable calculator (stub locally; UMA/omol on GPU when available) Streamlit UI with workflow page, real fast10_md.traj viz, and example LAMMPS shear data Tool surface designed for AWS Strands + Amazon Bedrock (not fully wired yet)
Challenges we ran into
Keeping claims scientifically honest (no “globally novel”) while still finding a useful gap Cognee tenant/auth and Bright Data SERP parsing quirks Multi‑GB production trajectories vs a responsive demo UI Separating demo viz from a real HPC path without overselling either Resisting a full LLM agent until tools and validators were stable
Accomplishments that we're proud of
A working end-to-end local loop: validate → stage → run → analyze → compare Real trajectory visualization + metrics from research data, not mocks Clear product story: Cognee = memory, Bright Data = web check, rules = safety Streamlit workflow that judges and demos can follow in minutes
What we learned
Closed-loop science agents win on constraints, not model size Deterministic gap logic is enough for an MVP; LLMs should orchestrate tools, not invent parameters Memory + live retrieval only matter if findings are structured and reusable Demo credibility comes from real trajectories + honest uncertainty, not flashy novelty claims
What's next for Mechanist
Write findings back into Cognee after every run Multi-experiment comparison and automated next-question proposals Harden analysis for production LAMMPS dumps at scale
Built With
- bedrock
- python
- streamlit