SSIA case study
Live demo · synthetic data

SSIA: Sustainability Strategy & Insights Agent

Pick a product. A deterministic engine computes its carbon footprint, and every row cites its emission factor. The strategy brief on top is LLM-written but never computes a number: every claim carries a citation, and a judge verdict is shown before you read it.

How this maps to a production AWS lakehouse
1 · Pick a product

The documented worked example: 0.5 kg aluminium + 0.3 kg plastic + 0.2 kg electronics. Reproduces the design doc's 9.4 kg CO₂e exactly.

2 · Configure the boundary
Cradle-to-gate
3 · The computed footprint (deterministic, no LLM involved)
9.40 kg CO₂e
±13% uncertainty from the data-quality mix · cradle-to-gate
emission factor × quantity, summed. Nothing generated.
Materials · process-based · 83.0%Materials · EEIO remainder · 17.0%
ComponentMethodComputationkg CO₂eFactor
Aluminium frameprocess0.5 kg × 12 kg CO₂e / kg6.00
Plastic panelprocess0.3 kg × 6 kg CO₂e / kg1.80
Electronicseeio0.2 kg × 8 kg CO₂e / kg1.60
How this was computed

Hotspot components use process-based factors (emission factor × mass). Non-hotspot components use EEIO sector factors to cover the long tail. Transport, when enabled, is mass × distance × mode factor.

The ±13% band is the emissions-weighted data-quality mix (process ±10%, EEIO ±30%, transport ±15%). Every row above cites the factor it used. Click any chip to inspect it.

A note on allocation: this demo allocates by mass for simplicity. In production PCF systems the allocation methodology (mass, economic, or product-movement-based) is the decision that moves the number, and it is only as good as the underlying data quality.

4 · The strategy brief (LLM-written · grounded · cited · judge-verified)
Decarbonization brief · pre-generated & cached for this demo

Aluminium drives this footprint; the electronics tail is the data gap.

The cradle-to-gate footprint is 9.40 kg CO₂e (±13%), quantified per the GHG Protocol Product Standard.

The aluminium frame is the largest single contributor; it carries the highest process factor in this bill of materials and the largest mass among hotspot components.

The electronics remainder is estimated with a sector-level EEIO factor, the least precise method in the mix; it widens the uncertainty band more per kilogram than any process-based row.

Recommended moves
  • Engage the aluminium supplier for supplier-specific emissions data first: it is the dominant row and currently computed on an industry-average process factor.
  • Request component-level activity data for the electronics assembly to move it from EEIO to process-based accounting and narrow the ±13% band.
  • Evaluate lower-carbon aluminium routes (recycled content, low-carbon smelting); any change flows through the same audited factor table, keeping the number defensible.
Trust panel
Judge verdictPASS
Citation coverage100%
Claims checked6
Numbers by LLM0

All 6 claims trace to provided evidence. No unsupported quantitative statements. Citation coverage 100%.

The one rule

The math is deterministic and auditable. The LLM narrates and strategizes over numbers that already exist; it never computes one.

A personal reconstruction of a concept I architected independently during my consulting tenure, built entirely with synthetic data and illustrative emission factors modeled on open sources (USEEIO, DEFRA-style). No client data, employer materials, or proprietary IP. Strategy briefs were generated with Claude against the computed evidence and cached; the citation-coverage check runs live in your browser.