Lean manufacturing analysis tool for real factory floor data. Load a project, run a full 12-section analysis — from ASME Flow Process Charts to ML-based bottleneck forecasting — and export a print-ready PDF report.
Built for: Python · Flask · Plotly · scikit-learn · Lean Manufacturing · Work & Method Study
Enter a workstation time study once (station names, cycle times, operator counts) and MechAI auto-derives a full 12-module factory efficiency report:
| # | Module | What it generates |
|---|---|---|
| 01 | AS-IS Summary | Standard time, normal time, takt time, bottleneck station, line/balance efficiency, NVA steps, units/shift |
| 02 | Flow Process Chart | ASME symbols (Operation, Transport, Inspection, Delay, Storage), zigzag flow, distance, lean recommendations |
| 03 | Line Balance Chart | Cycle time per station vs. takt line, bottleneck highlighted, idle time per station |
| 04 | Plant Layout Generator | 5 layout types — AS-IS (with backtracking), U-Shape, Process, Mixed, Static — interactive Plotly diagrams |
| 05 | Efficiency Comparison | Current vs. optimised across 4 charts (cycle time, KPIs, VA/NVA split, improvement summary) |
| 06 | SOP Generator | Standard Operating Procedure card per workstation — who/what/where/when, steps, safety, quality checks |
| 07 | Cost Calculator | Converts cycle time to ₹ cost; editable wage/worker/overhead; savings per unit/shift/day/month |
| 08 | Daily Production Report (DPR) | Editable shift-level table — qty planned/produced, dispatch time, EOD inventory |
| 09 | Shop Production Schedule | Gantt chart across shop sections for one full shift |
| 10 | ABC Inventory Analysis | A/B/C classification by annual value, Pareto chart, donut chart, replenishment strategy |
| 11 | Kanban Cards | Production pull-system cards for A/B items — part no., reorder point, max stock, supplier, bin location |
| 12 | ML Predictions | 4 scikit-learn regression models — Bottleneck Forecast, Output Forecast, Idle Time Prediction, Stockout Risk |
Every chart is interactive (Plotly) and the whole report exports to a single PDF via the browser's print pipeline.
- Backend: Flask 3.x, Python 3.11
- Data: pandas, NumPy, CSV-based project storage (no database)
- Charts: Plotly 5.x (server-generates figure JSON, client renders)
- ML: scikit-learn (LinearRegression, Ridge, Polynomial, MinMaxScaler) trained on station-level data at request time
- AI Chat: SmolLM2-360M-Instruct (HuggingFace), runs fully offline after a one-time ~720 MB download
- Frontend: Single-page vanilla HTML/CSS/JS (
templates/index.html), no build step, no framework
MechAI/
├── app.py # Flask routes — project CRUD, analysis endpoints, predictions, chat
├── analysis.py # All chart/report generation logic (Plotly figures, ABC/Kanban/cost calc, ML predictors)
├── insight_engine.py # Offline LLM chat engine (SmolLM2-360M) with jailbreak-resistant system prompt
├── launch.py # Standalone entry point — bootstraps the chat model, opens the browser, starts Flask
├── README.md # Project documentation
├── requirements.txt
├── projects/ # Saved project CSVs (one file per factory line)
│ └── Cooler_Manufacturer_Indore.csv
├── static/
│ └── logo.png
│ └── hero-bg.png
└── templates/
└── index.html # Full single-page UI (dashboard, new-project wizard, SOP/PDF modals, chat)
- Python 3.11+
- ~1 GB free disk space (for the offline chat model, downloaded on first run)
cd MechAI
pip install -r requirements.txtpython launch.pyThis starts a local Flask server at http://localhost:5000 and opens it in your default browser automatically. On first launch it also downloads the SmolLM2-360M chat model in the background (one-time, ~720 MB, requires internet); after that, MechAI Intelligence runs fully offline.
Alternatively, for development:
python app.py(runs Flask directly without the model-bootstrap/browser-launch wrapper, useful when iterating on the chat engine separately)
Projects are stored as flat CSV files in projects/, with three row types distinguished by a section column:
section,col1,col2,col3,col4,col5,col6,col7,col8,col9,col10,col11,col12,col13
META,Project_Name,Product_Name,Factory_Type,Location,ShiftMins,TargetOutput,ShiftsPerDay,WorkDaysMonth,PerfRating%,Allowance%,Wage/Hr,Workers,Overhead/Hr
STATION,#,Station_Name,Operators,Time(sec),Type,Component,Distance(ft)
COMPONENT,Name,UnitCost,AnnualUsage,LeadTime,Supplier,Location,ReorderQty,MaxStock- META (one row) — project metadata: shift length, target output, performance rating, allowance %, wage and overhead baselines.
- STATION (one row per workstation) — name, operator count, observed cycle time in seconds, type (
Operation/Inspection/Delay/Transport), linked component, and distance to next station. This is the only required input — everything else (NVA%, layout, schedule, SOPs) is auto-derived from it. - COMPONENT (one row per part, optional) — unlocks ABC Inventory Analysis and Kanban Cards when filled in.
New projects can be created through the in-app New Project wizard (3 steps: Setup → Workstations → Components) or by dropping a correctly-formatted CSV into projects/.
| Method | Route | Purpose |
|---|---|---|
| GET | / |
Serves the main single-page dashboard |
| GET | /api/projects |
List saved project CSVs |
| GET | /api/project/<filename> |
Project meta summary (station count, takt time, bottleneck, etc.) |
| GET | /api/project_data/<filename> |
Raw project CSV rows (META/STATION/COMPONENT) for editing |
| GET | /api/analysis/full/<filename> |
Runs all 12 modules and returns combined JSON (drives the main dashboard) |
| GET | /api/analysis/layout/<filename>/<type> |
Single layout diagram — type ∈ current, product, process, mixed, static |
| POST | /api/analysis/cost/<filename> |
Recalculates the Module 07 cost breakdown with live wage/worker/overhead inputs |
| POST | /api/predictions/<filename> |
Runs the 4 ML models for Module 12, accepts what-if scenario inputs |
| POST | /api/save_project |
Create/update a project from the New Project wizard |
| DELETE | /api/delete_project/<filename> |
Remove a project |
| POST | /api/import_csv |
Import a project from raw CSV content |
| GET | /api/model_status |
Chat model load state (ready / loading / not downloaded) |
| POST | /api/chat |
Ask MechAI Intelligence a question about the currently loaded factory data |
Four scikit-learn models are trained on-the-fly from the current project's station data:
- Bottleneck Forecast — sweeps target output upward and shows which station(s) cross the takt line first.
- Output Forecaster — predicts units/shift under what-if scenarios (added operators, performance improvement, station splitting).
- Idle Time Predictor — cumulative capacity wasted across a range of demand targets.
- Stockout Risk Scorer — MinMaxScaler-weighted risk score per component (cost 50% · lead time 35% · demand variability 15%), classified High/Medium/Low.
All four charts update live as the user adjusts demand target, operators-to-add, station-to-split, and demand-variability inputs.
An offline assistant (SmolLM2-360M-Instruct) embedded in the dashboard, scoped strictly to the loaded factory data via a hardened system prompt. It will:
- Discuss only the OEE/takt/cycle-time/bottleneck data from the currently selected project
- Refuse unrelated requests (general knowledge, code generation, creative writing, etc.)
- Interpret efficiency numbers honestly against fixed benchmarks (e.g. never call <50% line efficiency "acceptable")
- Resist prompt-injection attempts to override its scope
Quick-action buttons (Worst bottleneck?, Explain OEE, Takt compliance, ML prediction, Reduce cycle time, Full summary) are provided for common queries. Responses are capped at ~500 words (trimmed at the nearest sentence boundary) so answers stay concise without cutting off mid-sentence.
Click Export PDF (or Export Full Report PDF from the chat tab) to print the full analysis via the browser's native print dialog, in portrait orientation (default). The export includes the About, Dashboard, and Links tabs — the AI Chat tab is deliberately excluded from print (an empty/unused chat panel would otherwise waste pages). Print-specific CSS forces all chart backgrounds/colors to render correctly and ensures every plant layout type (Module 04) is included — not just the one currently active on screen.
- Demo project included:
Cooler_Manufacturer_Indore.csv— a 17-station desert air cooler assembly line based on real time-study field data. - The app is single-machine/local-first by design — no auth, no cloud sync, CSVs are the source of truth.
- Packaging with PyInstaller for a standalone desktop executable is planned but not yet wired into this codebase.
