Call prep & live notes

Google Cloud — Thu 2026-06-04, 10:00

Them: Lyly Willkens — BDR, Google Cloud Platform, DACH/EMEA. (Sales-dev, not an architect — she's the door to the right people, not the technical answer.)

Her agenda: Gemini Enterprise / Vertex AI · Cloud Run / Kubernetes · BigQuery. She asked us to bring technical challenges.

01What we want out of it

Pin the path from the $2k startup credits to the up-to-$350k tier — eligibility, timeline, whether it stacks with the trial credits.
Get her to book a Solutions Architect session on the hosting deploy — that's the real value a BDR can't give.
Extract a Vertex AI data-governance answer for pharma, in writing.
De-risk the Gen AI App Builder grounded layer (§ below).
Don't get sold GKE — we're on Cloud Run by design. Keep dev burning regcheck-pro credits.

02Questions to ask — capture the answer under each

Vertex AI data governance for pharma top

"If we move the engine from the Gemini API to Vertex AI — confirm (a) prompts/outputs not used for training, (b) EU-only residency (europe-west4), at-rest + in-transit, (c) a DPA / GDPR posture we can hand a Roche/Takeda procurement team." This is what converts a pharma buyer.

Gen AI App Builder credit — the grounded layer deal-breaker

Plan: thin grounded-generation + check-grounding in bring-your-own-retrieval mode on top of our own retrieval (we stay the citation authority).
① Deal-breaker: can grounded-generation be hard-constrained to cite ONLY our supplied passages, at clause-level fidelity — no synthesising beyond them? If not, the layer is dead.
② Credit scope: does the credit (vertex-genai-offer-2025) cover the grounding/generation APIs, not just Search queries? Does it touch our raw Gemini spend (which bills cash)? SKU list?
③ Confirm BYO-retrieval grounded generation works standalone (we will not ingest the corpus into a managed data store).

Vertex features for the engine

Context caching for the corpus (big cost lever) · grounding/citation (helps or duplicates ours?) · which Gemini tier for regulatory reasoning (accuracy > latency) · batch for the eval gold-set + daily corpus-diff digest.

Cloud Run cost at low-but-spiky traffic

Gateway is an auth front door — cold starts hurt the first request. Min-instances vs scale-to-zero economics? Best practice for Cloud Run + GCLB + Cloud Armor (we have all three)? Multi-region later? (Signals Cloud Run is deliberate; pre-empts the GKE pitch.)

Postgres — Supabase (cross-cloud) vs AlloyDB / Cloud SQL

Supabase is in AWS eu-central-1, compute in europe-west4. Price the latency + EU-residency + cost trade of moving to AlloyDB — don't commit, just price it.

BigQuery — only if it earns it

Candidate uses: usage-metering analytics + the reg-corpus-diff "digest" moat. Is BQ overkill vs Postgres at our scale; cheapest path to start?

03Asks to land before hanging up

The $350k startup-tier criteria + how to qualify + timeline.
A Solutions Architect working session on the hosting deploy (free, pre-launch).
Vertex AI DPA / data-residency confirmation for pharma — in writing.
#1: grounded-generation cites only our supplied passages at clause fidelity — in writing.
Gen AI App Builder credit SKU list + how to set SKU-scoped budget alerts so nothing bills cash.
Whether startup credits cover Vertex inference (not just infra).

04Hold the line on

Kubernetes / GKE — not needed; Cloud Run is the architecture. One sentence, move on.
No managed Vertex AI Search data store over the corpus — it uploads licence-restricted full text (breaks copyright tiering) + degrades clause-level citation. BYO-retrieval only.
PIF spec stays vendor-neutral — Preclari (product) runs on GCP; the standard does not "build on Google."
The Gemini key rotation is an internal item — don't surface it as a live problem.

05Notes & outcomes — relay these to me after; I'll fold them in

— to fill after the call —

06Reference — where Preclari is on GCP

Project preclari (europe-west4); dev/eval on regcheck-pro (trial credits). Hosting (PR #57 merged): Cloud Run ×3 — auth/metering gateway + MCP server (StreamableHTTP) + Zitadel; GCLB + Cloud Armor + managed TLS; mcp./login.preclari.com. Data: Supabase Postgres (txn pooler, eu-central-1) via Secret Manager; KMS holds ES256 signing keys. AI: engine calls Gemini via AI Studio today; not yet on Vertex; no BigQuery yet.

Also today: 1:00 PM marketing review with Fowler (Lederer). Andrew owns the materials + 5-slide deck; your partnership action is the MOU review.