Benjamin Arce
← Work
2025-08

Manufacturing Cost Intelligence System

A flight simulator for manufacturing decisions that replaces gut-feel spreadsheets.

manufacturinganalyticsstreamlit

Reality check: not an ERP replacement. But it’s already better than messy spreadsheets for multi-million-dollar decisions.

Quick scan
What it is An interactive what-if engine: change a cost driver (chips, labor, logistics) and see immediate portfolio P&L / margin impact, plus a short strategy brief that behaves like an analyst who actually read the numbers.
Why it matters Old workflow: someone asks “fuel is up, what happens?” then an analyst opens giant spreadsheets, changes cells, hopes formulas don’t break, and gives a partial answer days later.
Payoff Move a slider, see the impact, and have a strategy conversation instead of babysitting Excel. The system forces portfolio visibility and highlights hidden margin fragility.

What this actually is

Think of this as a flight simulator for manufacturing decisions.

A manager can move a single slider (“+15% on semiconductors”, “−5% logistics”, “+5% labor”) and see how the portfolio responds: gross profit, risky products, and levers that matter.

Then an AI writes a short strategy brief based on feasibility inputs instead of narrating charts.

Why I cared enough to build it

The real workflow was slow and fragile: multiple spreadsheets, formula landmines, and limited visibility beyond a couple of products.

It felt more like damage control than decision-making.

I wanted a system where you can change one driver, see portfolio impact immediately, and talk strategy like an adult.

What actually worked

  • The what-if engine: portfolio P&L simulation that makes impact visible in seconds.
  • The AI strategist: had to be trained to avoid chart narration; feasibility signals made recommendations realistic.
  • Portfolio health view: exposed high-volume products sitting on razor-thin margins (risk magnets).

What is still messy (being honest)

  • Static inputs (CSV-based). Great for controlled experiments, not production-grade without live integrations.
  • Feasibility scores are subjective; real deployments would need structured ops input.
  • Over-linear assumptions (no step-changes, discounts, or complex elasticity).
  • Simplified geography; a serious version needs region/plant-level costs, taxes, tariffs, risk profiles.

What I would do next

  • Live data integration (FX, fuel, freight index).
  • Multi-driver scenario stacking in one run (+10% steel, +5% labor, −3% logistics).
  • Monte Carlo forecasting for distributions (best/worst/bands).
  • Make AI interactive: compare interventions, ROI, implementation difficulty.

Technical details (plumbing)

  • Frontend: Streamlit
  • Compute: Pandas simulation engine
  • Visualization/export: Plotly + PDF export tooling
  • AI: strategy brief driven by explicit inputs (including feasibility) + hard constraints on what not to say
  • Architecture: core logic separated into /app/core; UI is just the face
  • Validation: startup checks to fail fast (unit mismatches, duplicates) instead of silently wrong output
← Back to Work