Which station and shift need attention?
1,075 kg scrapVA Line night shift, across 19 runs. Conversion was 92.5%, with 7.5 minutes average downtime per run.
FloorMind is a governed, read-only AI query layer for food manufacturers. Ask about yield, waste, orders, allergens or batch provenance in plain English. The default Ollama setup keeps model prompts on-site. If you configure a hosted provider, the question and relevant context are sent to that provider.
Your ERP and MES systems hold the answers. Getting them out is the problem.
Write SQL yourself (error-prone), wait for IT (slow), or export to Excel (outdated by the time it opens).
Anyone on the team asks in plain English. FloorMind routes the question to the right domain, generates and validates SQL, and explains the result.
For an initial demo, use synthetic data or an approved read-only copy. Live database access starts with a DBA-configured SELECT-only account, ideally on a replica. App validation adds another layer, it does not replace database permissions.
No plugins into your ERP. No data pipelines to build. One read-only connection is enough.
Your question is mapped to one of six business domains (production, traceability, orders, compliance, staff, stock) so the model only ever sees the tables that matter.
17 common factory questions return pre-tested SQL without an LLM call. Other questions use the local model by default, or a hosted provider if configured.
Generated queries pass sql-sop analysis and FloorMind's app validator. A SELECT-only database account remains the enforcement boundary if a query gets past those checks.
Results render as a table or chart with a plain-English explanation, and the full query record lands in a tamper-evident audit ledger.
Captured from the running demo. No mockups.
Complex questions across production, sales and traceability. SQL is generated, safety-checked, executed and audit-logged.
VA Line night shift, across 19 runs. Conversion was 92.5%, with 7.5 minutes average downtime per run.
Morrisons Fish Pie Mix, from 1,428 kg delivered at a weighted GBP 15.69 per kg.
North Atlantic, wild-caught, Iceland. Linked to one trace ID in this sample dataset.
These examples use synthetic demo data and a hosted Groq model. No customer or production data is shown.
Start with synthetic data on your machine. Pre-built queries work without an API key; install Ollama for open-ended local questions. A hosted provider is optional and receives the prompts and context needed to answer.
# 1. Get the code and set up Python 3.11+
git clone https://github.com/Pawansingh3889/FloorMind.git
cd FloorMind
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# 2. Seed 60 days of synthetic factory data
python scripts/seed_demo_db.py
# 3. Run the app
streamlit run app.py # โ http://localhost:8501
# Optional: local LLM for open-ended questions
ollama pull gemma3:12b
# Optional: kiosk dashboard for a TV screen
FLOORMIND_KIOSK=1 streamlit run kiosk.py
Full setup guide with Mac and Linux notes: docs/setup.md
The software runs on your premises. The subscription covers everything that keeps it working there.
We deploy FloorMind on your hardware and connect it to your read-only replica: SQL Server or PostgreSQL, mapped to your schema.
Your tables, your terminology, your domain rules. We tune the question router and the pre-built query library to your operation.
Priority fixes, model and dependency updates, and a named engineer who knows your installation.
Every query linted, logged, and hash-chained. Governance documentation written for corporate IT and BRC auditors, not just developers.
Excluding VAT. One fee per site, with no per-seat charge. Choose the support and deployment level that fits your operation.
Single site, self-serve install, evaluation and small teams.
excl. VAT, billed monthly
30-day pilot. Annual billing saves 20%.
A running production site with a live database and named recipients for alerts.
excl. VAT, billed monthly
30-day pilot available. Annual billing saves 20%.
Multi-site groups, custom integrations, procurement and auditor requirements.
excl. VAT, annual agreement
Annual agreement, priced per deployment.
FloorMind was designed so a security reviewer can verify the controls themselves. The full model is in docs/architecture.md.
| Layer | Control | What it prevents |
|---|---|---|
| L1Database grants | DBA-configured SELECT-only user with explicit DENY on writes | Any write reaching the source database |
| L2Connection routing | During setup, point FLOORMIND_DB at the approved read-only replica | Any connection to the production primary |
| L3Static analysis | Rule-based sql-sop lint before query execution; current source and package activity are linked below | Unsafe SQL patterns and poor query constructs |
| L4App validator | Second independent parse with row caps | Defence in depth and unbounded result sets |
Open source by Pawan Singh Kapkoti, with development in the Governed Agent Stack. Source, tests and package activity are linked below; companion tools are distinguished from controls used by FloorMind.
In-process rule-based SQL lint, called by FloorMind's validator before a query executes.
Built into this app: read-only statement checks, injection-pattern checks, table validation, row caps and the sql-sop pass.
FloorMind's audit log can mirror events to this append-only, hash-chained ledger when installed and enabled. It records actions; it does not approve queries.
FloorMind installs it as an open-source dependency. The ledger can be verified independently.
Exposes the linter to MCP-compatible clients. It is not part of FloorMind's Streamlit query path.
Separate read-only SQL MCP server with driver-level restrictions, SQL AST validation and sql-sop lint. Optional role checks can be added with query-warden.
Optional role-based policy check for allowed SQL tables and columns. It is not currently wired into FloorMind.
These PePy charts include CI downloads. They show package activity, not unique users or an independent security audit. PePy version metadata can lag PyPI; use each project source or PyPI page for current release details.
Start with synthetic data or an approved read-only copy on your machine. Before a production connection, your DBA configures a SELECT-only account. We can answer a question your team asks every week.