1 / 8
←→ navigate
01

CSF Kenya × Google Cloud

We moved our AI platform onto GCP — now we need two things to scale it

Gemini
Enterprise + Antigravity
Compute
Engine Quota
🇰🇪 Complementary Schools Foundation Kenya  ·  🇬🇧 The Children's Education Foundation (UK Charity No. 1207489)

Felix Edwards & Mulili (CSF) · with Amit Sarkar (Google Cloud, EMEA)

02
Who We Are

CSF Kenya & The Children's Education Foundation

The Kenyan arm of a UK education charity

CSF Kenya is the Kenyan counterpart of The Children's Education Foundation (CEF) — a UK-registered charity (Charity No. 1207489).

Our Mission

The Children's Education Foundation is dedicated to empowering complementary schools in low-income areas. We provide sustainable, cost-effective, quality education by enhancing their financial viability and improving leadership and teaching practices — driving positive outcomes for pupils and a brighter future for children in our communities.

🏘️ The need~36% of Nairobi lives in informal settlements; ~60% of children there attend LCPS, which the state does not fund.
💸 The problemLCPS are financially fragile — fees of $20–$50/term, collection rates of only 30–60%, teachers paid late.
03
What We Do

Modular Support & Proven Impact

Increase revenue, reduce cost, put the school at the heart of the community

Depending on each school's needs, we provide modules of support — then step back so the school can grow on its own:

🖥️ School Management SystemSoftware to track fees & records — no upfront cost
☀️ Solar + batteriesCuts electricity cost
💧 Water tanksCuts water cost ~50%
🔥 Efficient stovesCuts firewood cost ~85%, less smoke
🔬 Labs & booksComputer/science labs, textbooks, reading culture
🎓 TrainingLeadership & teacher training, soft skills
Pilot — Elite Heritage School, Kayole

411 pupils, fees ~$50/term. With the school management system, fee collection rose from an estimated 50–60% to 92% — enabling teachers to be paid on time & in full, new teaching materials, and facility upgrades. A far more resilient institution.

The ambition: scale to 50+ schools across Nairobi & Mombasa, partnering with organisations like WERK & Dignitas.
04
Our Platform

Lovelace — Our AI, Now Running on GCP

Migrated from AWS → Google Cloud on 3 Oct 2026 · now live at 136.70.218.16

To deliver that support at scale, we built Lovelace — an AI field-operations assistant for CSF staff and schools. It is not a single model, but a mixture of distilled Qwen 3.8 27B models, each specialised for a task:

📝 Reports modelSchool impact, KPI & progress reports, meeting-note generation
💻 Go (Golang) modelBuilds & maintains the backend services that run the platform
⚛️ React modelBuilds the admin dashboard & the csfkenya.com school websites
🎙️ Speech & OCRTranscribes meetings/voice notes, reads school documents into data
Already live on GCP

39 schools onboarded · a Telegram bot with 30+ tools · weekly Gemini TTS audio meeting summaries · a Google-Meet note-taker · dashboard & Telegram mini-apps — all on a single Compute Engine instance (136.70.218.16, Ubuntu 24.04).

✅ We have already committed to Google Cloud — the full stack now runs on GCP, and we use Gemini for TTS and routing today. The distilled model weights run on-prem on an NVIDIA DGX Spark, with Gemini as the cloud fallback when needed.
05
The Ask · 1 of 3

Gemini Enterprise + Antigravity

So Lovelace can write code & fix bugs on its own

We want Lovelace to write code, ship features and fix bugs itself — using the Antigravity CLI (agy) via Gemini Enterprise — especially in Go and React, for the moments when Mulili (our only developer) can't get to it.

🧩 Bundled accessAntigravity enterprise access is included in Gemini Enterprise seats — SSO, pooled quotas, budget caps, audit logging. No separate licence.
🤖 Agentic codingAgents plan, write code, run terminal commands, debug, and verify in a sandbox — orchestrated via the CLI, SDK or IDE.
🔒 Enterprise safeCode & prompts never train base models; isolated execution sandboxes; central audit logs on GCP.
🧠 Multi-modelNative MCP support — Antigravity agents can call our own distilled models alongside Gemini.
Standard (~$30/seat/mo — no-code agents, data connectors) vs Plus (~$50–60/seat/mo — code-first ADK agents, higher quotas, VPC-SC/CMEK). Indicative list price — to confirm with Amit.
06
The Ask · 2 of 3

Compute Engine vCPU Quota

Headroom to grow the agent as more schools come online

Lovelace runs today on a single Compute Engine instance (136.70.218.16). As schools onboard and the agent takes on more concurrent work, it needs more cores.

8
vCPUs today · c2-standard-8
→
scale up
32
vCPUs target · c2-standard-32
What we're requesting

Raise the C2 CPU quota from 8 → 32 (c2-standard-32, 32 vCPU / 128 GiB) in two regions, plus GPU access for model inference:

  • 🇺🇸 us-east4 (N. Virginia) — where the instance runs today
  • 🇬🇧 europe-west2 (London, UK) — closer to our UK charity base & for redundancy
Quota is per-region, so we need the increase approved in both us-east4 and europe-west2.
07
The Ask · 3 of 3

Vertex AI — TTS Quota & Argon Fallback

Lovelace thinks on the DGX Spark — Gemini is the cloud safety net

TTS quota

⚡ Raise RPM / TPM / RPD on gemini-flash-tts. Today the Tier-1 ceiling (10 RPM / 100 RPD) throttles our weekly audio meeting broadcasts to ~2 runs/day. Provisioned Throughput would remove the 429 rate-limit wall for real-time audio.

Gemini 4 (Argon) as LLM fallback

🧠 Lovelace's distilled weights run on-prem on the NVIDIA DGX Spark — the agent "thinks" locally. But we need to offload LLM compute to Gemini 4 / Argon on Vertex AI when:

  • the Spark models misbehave and we need a reliable fallback, or
  • Lovelace is doing many things at once — writing reports, coordinating field visits, and speaking in meetings simultaneously.

We're requesting access + quota for Gemini 4 / Argon as that cloud safety net.

08

Next Steps

"Enable the platform, and CSF can scale from one pilot school to a whole network."

  • 1️⃣ Gemini Enterprise — a trial / seats for the CSF dev team, with Antigravity CLI access for autonomous Go & React development.
  • 2️⃣ Compute Engine quota — C2 CPUs 8 → 32 in us-east4 & europe-west2, plus GPU access.
  • 3️⃣ Vertex AI quota — Gemini TTS RPM/TPM/RPD increase, plus access to Gemini 4 / Argon as an LLM fallback.
  • 4️⃣ A technical point of contact — a Google SE/CE to help right-size quota and the Gemini Enterprise edition.
  • 5️⃣ Nonprofit pathways — any Google for Nonprofits / credits applicable to a registered education charity.

Thank you, Amit 🙏

Felix Edwards & Mulili · CSF Kenya  ·  [email protected]