BRAHM

Bi-directional Research & Analysis Hub for Multimodal Science

A personal project — a self-hosted, multi-agent AI platform for automating materials science research workflows. Built to reduce the time between a scientific idea and a validated result.

What It Does

Modern research is slowed by information overload, fragmented knowledge, repetitive analysis, and documentation overhead. BRAHM automates the mechanical parts of the research process while keeping the researcher in control of every scientific decision.

The pipeline moves from literature to knowledge to insight to communication through a unified, evidence-driven workflow.

Related MCP server: tooluniverse

Architecture

BRAHM is built as an ecosystem of specialised agents, each responsible for one layer of the research process.

Literature Discovery → Knowledge Extraction → Gap Analysis → Simulation → Documentation
      SHANI                  SHANI              SHANI       Vishwakarma      GANESH
                                          Chitragupta

Agents

Pipeline Stages

SHANI — Literature Pipeline

GANESH — Document Pipeline

Tech Stack

• Runtime: Python, FastAPI, SQLite, FAISS

• LLMs: Groq, Gemini, Cerebras (cloud) + Ollama (local, for sensitive data)

• Computation: Quantum ESPRESSO 7.5 for DFT

• Embeddings: all-MiniLM-L6-v2 for vector search

• Infrastructure: WSL2 on Windows 11, self-hosted

Design Principles

Human judgment remains central. BRAHM assists research. Researchers direct research. Scientific decisions always belong to the researcher.

Scientific integrity above speed. Evidence is more important than confidence. Uncertainty is surfaced rather than hidden.

Privacy by design. Instrument data and DFT structures never leave the local machine. Only published paper text reaches cloud APIs.

LLM-agnostic. Each agent uses the best available model for its task. The orchestration layer is model-independent.

Status

Active development. Core pipeline (S1→S5) is operational. Document generation (GANESH G1→G5) is functional. S5.5 finding reconstruction and full grounded document generation are in progress.

Known limits

Stated plainly, so the system is not read as further along than it is.

• Grounded writing has not met its bar. The target is ≥90% of cited claims supported by the cited source, with 0 fabricated. The first 50-claim audit measured 86% supported, 0 fabricated — the fabrication half is met, the support half is not. The remaining failures are correctly-copied numbers attached to the wrong subject, which no threshold change reaches.

• 31 of the 64 MCP tools have never been exercised end to end. They are the destructive, long-running and costly ones; there is no safe harness for them yet.

• Chitragupta has two persistence layers with overlapping purpose — brahm_db/ → brahm.db (live) and api/routers/store.py → brahm_knowledge.db (unused). Consolidation is pending.

• VIDUR does not yet parse XPS, gas sensing or PL, and has no peak finding, phase ID, Scherrer or Tauc. Unsupported techniques are named explicitly so they surface as a question rather than a confident wrong guess.

• Quantum ESPRESSO binaries and pseudopotentials live outside the repo; QE_BIN_DIR, QE_PSEUDO_DIR and VISHWAKARMA_WORKDIR must be set for Vishwakarma to run.

Project Structure

brahm/
├── agents/
│   ├── shani/          # Literature pipeline
│   ├── chitragupta/    # Knowledge custodian
│   ├── ganesh/         # Document generation
│   ├── vidur/          # Characterisation data preparation
│   └── vishwakarma/    # DFT computation
├── brahm/              # Shared registry and utilities
├── brahm_dashboard.py  # Service health and control UI
└── mcp_server.py       # MCP entry point

Built by Shardul Khanduri — MSc Physics, materials science and AI systems.