[api] AIN-299 W5: public live counter (aggregate-only)

[hizrianraz]

GET /v1/stats/public/counter — 4 aggregate fields, 60s cache, NO PII (3 unit tests enforce). AIN-293 Slack code deferred to post-Spark Fri.

---

Note

Medium Risk
New unauthenticated public endpoint reads production aggregates (routing, agents, cost); privacy is schema- and test-guarded but cache is per-process only until multi-worker scaling.

Overview
Adds a public GET /v1/stats/public/counter endpoint on the existing stats router so the marketing site can poll aggregate-only live metrics: routed decisions (lifetime and 24h, cheapest_clearing_floor on routing_outcomes), active agents, and cumulative succeeded inference cost in USD.

Responses are built from a single multi-subquery SQL read, wrapped in a CounterResponse schema with fetched_at and cache_ttl_sec. A 60-second in-process cache limits DB load under frequent polling; module notes call out single-worker behavior and a future Redis/shared cache if the API scales out.

Privacy is explicit in the API contract (no agent/tenant/per-row IDs). tests/unit/test_audit_counter_no_pii.py locks schema fields, types, and absence of UUID-shaped strings in JSON. The OpenAPI v1 contract smoke test now expects this route.

^{Reviewed by Cursor Bugbot for commit 13805df. Bugbot is set up for automated code reviews on this repo. Configure here.}

[linear-code]

AIN-299 [Marketing] Live public audit counter — ainfera.ai transparency surface (eat-our-own-cooking)

Live public audit counter — eat-our-own-cooking transparency surface

Polsia-style transparency move on Ainfera's terrain. Free positioning, single Aulë day, no critical path.

What

Component on ainfera.ai (footer or hero) that polls /v1/audit/public every N seconds (60s reasonable) and displays live aggregate:

Metric	Source
Routed decisions (lifetime)	count(audit_events WHERE event_type='inference.routed')
Agents served	count(distinct agents.id WHERE status='active')
Cumulative inference cost routed	sum(inferences.cost_usd)
(Optional) cumulative savings vs baseline	from routing_outcomes delta corpus once Labs daily cron has 14d data

Display in dark mode (Brand v1.3: #070B14 bg · #E8EDF5 ink · #4D95E8 accent · IBM Plex Sans 500).

Why

Perplexity external pressure-test 2026-05-28 surfaced this as the highest-extraction insight: Ainfera's /v1/audit/public feed is live but no marketing surface exposes it. Polsia's live dashboard of 7,600+ companies is a marketing wedge — Ainfera's equivalent is the audit chain. Same play, different proof artifact.

Eat-our-own-cooking corollary: routing 100% of internal LLM calls (fleet + Aulë IDE) through ainfera-inference is already locked. Surfacing the audit publicly = marketing for free.

Scope

• Add /api/audit-counter aggregator endpoint in ainfera-ai/api (60s cache; tenant-aggregate counts, no per-agent leak)
• Marketing site component (ainfera-ai/web or current marketing repo): React/Vue/static, polls endpoint, animates count-up
• Brand v1.3 styling
• Verify no PII leaks via audit-counter (only aggregate counts, never agent_id or tenant_id)
• A11y: numbers announced for screen readers only on full minute (not every tick)

Acceptance

• Counter renders on ainfera.ai
• Polls /v1/audit/public aggregate every 60s
• Numbers monotonically non-decreasing across reload (proves not fake)
• 0 PII in network response (verify in browser devtools)
• Dark-mode compliant
• Smoke test: counter increments after a fresh routed call

Apply window

Post-Spark migration (Week of 2026-06-08+). Not on critical path. Single Aulë day once daily cadence is stable.

Linked

• Memory build(deps): Bump astral-sh/setup-uv from 5 to 7 #1 — eat-our-own-cooking thesis
• AIN-296 (Aulë IDE wire — feeds the counter at fleet-scale)
• AIN-298 (Daily training cadence — feeds the "savings" metric once 14d data accrues)

Source

External pressure-test 2026-05-28 (Perplexity analysis) — sole actionable extraction. Other Perplexity recommendations (local Qwen judge, α=25%, agentium.yaml) rejected for lock conflicts.

Review in Linear