Capstone Project · 2026
GreenScale
A carbon, cold-start, latency and cost aware routing plane for multi-region serverless.
presented by
Anshuman Mohanty
B.Tech CSE · Cloud Computing · GF202217744
mentor: Mr. Ashish · Shoolini University
live demo at greenscale.run
01
where we are today
Serverless routing is one-dimensional.
Every production load balancer, from Google Cloud LB to AWS Latency-Based Routing to Cloudflare Argo, picks the closest region by round-trip time. Three other variables that move the cost and externalities of a request by an order of magnitude are invisible:
- Grid carbon intensity — varies 5×–10× across regions any moment of the day.
- Cold-start probability — a cold region adds 200–1500 ms before the first byte.
- Hourly cost — region pricing differs 10–15%; non-trivial at SME scale.
A routing system that ignores three of its four most important inputs is not an optimum. It is a default.
02
what the latest papers do — and don't do
Two specific gaps remain open.
Recent state of the art (2025–2026)
- LACE-RL (Feb 2026) — RL keep-alive tuning. Single region.
- Aceso (Mar 2026) — carbon-aware placement. Design-time.
- Spatio-Temporal Shifting (Dec 2025). Batch only.
- Green-by-Design (Feb 2026). Constraint solver, not runtime.
Open gaps
- Gap 1. No per-request, multi-region, runtime routing layer that jointly optimizes carbon, cold-start, latency and cost.
- Gap 2. Existing solutions are research prototypes — not deployable, transparent middleware that an SME can adopt without rebuilding the platform.
03
the proposal
A transparent routing plane in front of any serverless deployment.
For every request, GreenScale computes a per-region score:
score(r) = w_lat·latency(r) + w_carbon·CI(r) + w_cold·P_cold(r) + w_cost·cost(r)
chosen = argminr score(r)
- Linear & transparent — every weight, every score is logged. Auditable for CSRD and India BRSR.
- Provider-agnostic — runs as a Cloud Run service in front of any HTTP serverless backend. No platform changes.
- Re-tunable at request time — operators can shift the carbon/latency trade-off live, no deploy.
04
architecture
Three Cloud Run services. Zero idle cost.
┌──────────────────────────┐
│ User browser │
└────────────┬─────────────┘
│ HTTPS
▼
┌──────────────────────────┐
/, /pitch, /report ◄─┤ greenscale-frontend │ nginx static
│ Cloud Run · min=0 │
└────────────┬─────────────┘
/api/* reverse-proxy
▼
┌──────────────────────────┐
/v1/route │ greenscale-backend │ FastAPI · 200-line scorer
/v1/regions │ Cloud Run · min=0 │
/v1/decisions └────────────┬─────────────┘
▼
┌──────────────────────────┐
POST /decisions │ greenscale-db │ SQLite + GCS round-trip
GET /decisions │ Cloud Run · min=0 │ RPO ≤ 30s
└────────────┬─────────────┘
▼
gs://greenscale-state
05
the result
Up to 87% less carbon. Under 50ms median overhead.
500-request synthetic benchmark, 12 user locations across 3 continents, full 24-hour UTC carbon cycle.
87%
peak operational carbon reduction vs latency-only routing (worst-case region)
38ms
mean latency overhead. P95 = 142 ms.
67%
reroute rate — chose a greener region than latency-only would.
₹0
idle cost. min-instances=0 across all three services.
~80ms
end-to-end warm round trip.
200
lines of routing code. Zero ML black box.
06
vs. the published state of the art
Complementary, not competitive.
| property |
LACE-RL |
Aceso |
Green-by-Design |
GreenScale |
| per-request | — | — | — | ✓ |
| multi-region | — | partial | — | ✓ |
| runtime (not design-time) | ✓ | — | — | ✓ |
| transparent / auditable | RL black-box | heuristic | constraint | linear, logged |
| drop-in (no platform mod) | — | — | — | ✓ |
| all four signals | cold + carbon | carbon + cost | carbon + constraints | ✓ all four |
LACE-RL inside each region + GreenScale across regions is the right production stack.
07
why now, why for SMEs
CSRD & India BRSR are turning carbon into a compliance line.
2025
India SEBI mandates BRSR Core for top 1000 listed companies. Scope-2 cloud carbon is in scope.
2025
EU Corporate Sustainability Reporting Directive expands to ~50 000 firms. Scope-2 reporting required.
2026+
SMEs in supplier chains will inherit the requirement. Most run on managed serverless and have no path to adopt platform-internal carbon schedulers.
GreenScale is the drop-in path. Three Cloud Run services, install in 10 minutes, full audit trail, no platform changes.
08
what's next
From capstone to product.
- Live carbon API. ElectricityMap or WattTime per-region 5-minute resolution.
- SDK. @greenscale.route() decorator on existing Cloud Run handlers — adoption in two lines.
- Hosted control plane. Multi-tenant, billed per million decisions.
- Compliance exports. CSRD + BRSR-aligned CSV from the audit table.
- Multi-cloud. AWS Lambda & Azure Container Apps adapters.
Capstone today · production-grade middleware in 9 months · the default sustainability layer for managed serverless.
09
thank you
Questions?
Live demo
greenscale — on Cloud Run
/ for dashboard · /pitch for these slides · /report for the full report
Anshuman Mohanty
B.Tech CSE · Cloud Computing
GF202217744 · mentor Mr. Ashish
Shoolini University · Solan, H.P.
10