Architect Toolkit

GenAI architecture isn't a technology problem — it's a judgment problem. This reel hands you the working architect's kit: selection matrices, NFR checklists, cost models, and the gates between POC and production.

▶ Watch this reel

What you'll learn

Choosing & justifying
Architecture patterns
NFRs & cost model
POC to production

Remember this

Select use-cases on value x feasibility — kill weak ones in the meeting; high-value + low-feasibility needs a human checkpoint first
Every GenAI system composes five blocks — model, retrieval, orchestration, guardrails, observability; buy undifferentiated blocks, build the one you win on
GenAI's NFR triangle: accuracy SLO (evals in CI), latency (TTFT + streaming), cost per completed task — optimize one, budget the other two
POC to production runs through gates — eval, safety, scale, operate — each with an owner; demo failure modes on purpose to earn trust

Use-case selection

Axes: value (measurable baseline, owner, annualized impact) x feasibility (verifiability, error tolerance + human loop, data readiness, latency fit).
High value + low feasibility → don't kill, add a human checkpoint and re-score.
Kill low-value on the record — it buys credibility for the real 'yes'.

Reference architecture

Five blocks: model endpoint · retrieval · orchestration · guardrails · observability. Build-vs-buy per block: undifferentiated → buy; standard-needs-assembly → configure; differentiating → build. Record reasons.

NFRs (GenAI-specific)

Accuracy: SLO on a golden eval set (AG-21); evals in CI.
Latency: TTFT + tokens/sec are user-visible (GA-07); streaming is standard; budget per interaction type.
Cost: per completed task = LLM calls + retrieval + guardrails + infra; compare to human baseline.
Triangle: improve one, budget the other two explicitly.

POC → production gates

G1 eval · G2 safety · G3 scale · G4 operate — each with an OWNER. Risk register: likelihood x impact → named mitigation → cites the reel that teaches it. Client communication: demo the failure modes on purpose. Confidence is the deliverable.

Cross-links

GA-01 (probabilistic nature → feasibility), GA-07 (latency anatomy), GA-09/10 (guardrails + RAG), GA-29 (buy options), AG-21/22 (evals, red-team), AG-23 (cost discipline). AG-25 reuses this whole kit for agents.

Code: The architect's one-page scorecard

USE-CASE      value  feas.  decision   owner
support-triage  11     17     POC ✓      VP Support
contract-draft  14     11     +human HITL Legal
auto-approvals  13      6     KILL       (recorded)

GATES:  G1 eval ✓  G2 safety ✓  G3 scale ✓  G4 operate ✓
NFRs:   accuracy ≥ 92% on golden set (CI-gated)
        TTFT < 800ms, streaming on
        cost/task ≤ $0.08 @ 40k tasks/mo
RISKS:  5 open, 4 mitigated — each cites a reel

One page. Defensible. Signed.