Min ticket$1M+
Time to close6+ mo
RecourseFull
GeographyOECD
Need salvage to lend. Retrofit equipment resells at ~10%, so they require full corporate recourse on top — which only investment-grade balance sheets can post. Lights off below $1M.
Physics-informed underwritingCalibrated uncertainty bounds disclosed per project
Capital that meters itself.
Non-recourse loans for industrial efficiency retrofits, underwritten by physics-informed AI and verified by IoT meters. Capital from institutional LPs worldwide flows on-chain to industrial SMEs across India and Southeast Asia, repaid from the energy savings the upgrade generates.
One meter. Three views.
Lenders find yield. Borrowers find capital. Builders find every primitive auditable.
01 · Lenders · LPs
Earn 10–14% yield on industrial efficiency credit.
Non-recourse loans sized to the P5 floor of a calibrated 90% PI. DSCR @ P5 ≥ 1.30× hard covenant. Monthly USDC.
02 · MSME borrowersUpgrade your factory. Repay from the savings.
4–6 weeks to close. Non-recourse to your business. ₹20L–₹100Cr facility sizes for vetted Indian MSMEs.
03 · BuildersEvery primitive auditable. Every prediction reproducible.
Six primitives, one ledger. Vault custody on RWA rails; calibrated underwriting + IoT M&V + audit-hash commits.
We make energy savings a bankable asset class.
kWh × $/kWh = real savings
How it works.
One deal, end-to-end: audit, capital, retrofit, repayment, yield.
Banks need salvage. ESCOs need guarantees. We underwrite the savings.
Ticket-size coverageFilled portion = the ticket sizes each can actually serve.
AmperisBanksESCOs
Min ticket$5–10M+
Time to close12–24 mo
StructureESPC guarantee
GeographyUS + EU
Need a corporate guarantee. The 1980s answer: don't underwrite the savings — wrap them in the borrower's creditworthiness. Only investment-grade enterprises qualify.
Min ticket$25K
Time to close4–6 wk
RecourseNone
GeographyIndia · SEA
Underwrite the savings directly. Calibrated PINN + IoT M&V + legal assignment of the kWh delta. No salvage needed, no corporate guarantee needed — 10× the addressable market.
Underwriting you can verify.
LOO R² of +0.56 on 72 KISEM industrial audits, after pretraining on 14k US-DOE IAC rows. The 90% prediction interval is distribution-free.
| Method | R² (LOO) | MAPE-median | 90% CI coverage | Verdict | Notes |
|---|---|---|---|---|---|
| Sector-median ratio (heuristic) | 0.16 | 41.7% | — | Decent | What unaided underwriters do today |
| BEE physics formula (industry default) | 0.50 | 37.5% | — | Good | Leakage-known compressed-air only |
| Amperis CatBoost | +0.28 | 44.7% | ±67,679 kWh (split-conformal) | Decent | IAC pretrain + KISEM finetune, deployed |
| Amperis TabPFN (this product) | +0.56 | 41.6% | ±69,254 kWh (split-conformal) | Good — within SOTA band | TabPFN in-context on IAC + KISEM. Hollmann et al., Nature 2025. |
Caveat for lendersIAC's "realized savings" is client-reported at 6–9 month phone follow-up, not metered M&V. Our own metered M&V loop closes the gap post-deployment via IPMVP Option B telemetry.
What R², MAPE, and 90% CI mean
R² — how much variation the model explains
- +1.0 perfect · 0.0 guessing the mean · negative actively misleading
- Industrial energy-savings papers report 0.5–0.7. Our TabPFN sits inside that band with ~50× less data.
MAPE-median — typical % error per prediction
- 10–20% excellent · 20–40% good for lender debt-sizing with a P5 floor · >60% don't underwrite
- At 41.6%, a 100k-kWh prediction lands in 60–140k typically. Why we underwrite to P5, not P50.
90% CI coverage — is the band honest?
- Split-conformal (MAPIE 1.4) gives a distribution-free statistical guarantee of coverage.
- Not an estimate — a mathematical proof under exchangeability. The reason lenders size debt against P5.
Why TabPFN is the headline, not the serving model
Two models, two roles
- TabPFN — benchmark. R²=+0.56 LOO on a 6-feature corpus. Proves a foundation-model generalises out of a tiny Indian audit set when conditioned on 14k IAC rows.
- PINN unified — what serves live underwriting. Ingests all 21 audit fields (leakage_pct, rated_kw, motor count, plant context). On small ECMs it delivers a lender-grade band where 6-feature TabPFN would floor at zero.
Next iteration
Retrain TabPFN on the 21-feature audit schema, validate bands no longer floor at zero, then flip the serving default. The headline number is preserved; the audit-signal moat becomes part of the model at both benchmark and serve time.
Today: every project page reads "Underwritten by PINN unified" — that's the model sizing the loan. The TabPFN R²=+0.56 above is the LOO score on the corpus it was trained on, not a live-serving claim.
How to verify these numbersEvery seed deal on /projects publishes its P5 / P50 / P95 band with the same model and methodology. The full sizing math is in the Underwriting Policy. A public Colab notebook running the LOO-CV reproduction on the IAC subset ships with V1 mainnet.
Hard to copy.
Five layers, sorted by replication difficulty.
A $500K compressed-air retrofit resells for ~$50K post-default — about 10% recovery. The same $500K of GPUs resells for ~$300K — about 60%. Retrofit equipment is custom-fitted to a specific factory's pipework; uninstallation often costs more than resale; the secondary market for used VFDs, chillers, and heat-exchangers is thin. A non-recourse loan needs two legs: salvage value (sell the asset) and cash flow (income the asset generates). Banks rely on the salvage leg, so they won't touch retrofits. We have only the cash-flow leg — but it's a leg only a calibrated physics model can build. Anyone with a balance sheet can compete in GPU credit. Nobody can compete here without our underwriting tech.
Answers.
How yield is sized, when distributions arrive, exit liquidity.