Early access — now onboarding teams
Pong sits between your company and every LLM provider. Every query is policy-checked, enriched with your organization’s memory, routed to the best allowed model, and recorded in an immutable audit log.
The problem
Four market realities, one structural gap: the layer between your company and the model providers doesn’t exist yet — unless you build it yourself.
Enterprise AI spend hit $37B in 2025 — 3.2× year over year — and token bills keep doubling. Uber blew through its annual AI budget in four months and imposed per-employee caps. Tesla capped employee AI spend at $200 a week. AI became the fastest-growing uncontrolled line item in the company.
93% of CIOs say different models win different tasks, and 81% expect to run two or more LLM providers. But routing between them breaks the thing that matters most: memory, context, and history don’t port across models. Every switch starts your organization back at zero.
Courts have compelled AI providers to preserve user conversations — only Zero Data Retention API customers were exempt. What your provider keeps is no longer a settings toggle; it is discoverable evidence. Retention terms now belong in your compliance posture, attested per request.
DoorDash, Coinbase, Uber, Walmart, Goldman — every large company built an internal LLM gateway to get routing, governance, and cost control. Everyone else can’t staff a platform team to build and maintain one. Pong is that gateway, delivered as a product.
The insight
Zero Data Retention and persistent memory are mutually exclusive — unless a layer you control holds the memory.
Providers can retain nothing while your organization still remembers everything, because the memory lives in your governed store — portable across every model. Pong is the memory the model providers can’t hold, and the audit trail regulators demand.
How it works
Every request — from any team, to any provider — moves through the same governed pipeline. Nothing is routed on price until it has passed policy.
Compliance eliminates models before economics ranks them. Department allowlists, ZDR requirements, budget state, and residency rules filter the model pool first — cost optimization only ever chooses among models you are allowed to use.
Every request is classified for task type, sensitivity — with PII, secrets, and financial-data detection — complexity, and consequence of failure. That classification drives every decision downstream.
Hierarchical memory — organization, department, user — is retrieved before model choice, filtered by ACLs. Cheaper models get the context that made frontier models good, which is what makes cost-optimized routing safe.
Consequence-weighted selection: the cheapest eligible model that meets the quality bar wins. High-consequence work is never silently downgraded to save money.
Per-task scaffolding — prompt structure, tools, verification — tuned to the job. The same model in a better harness routinely delivers about 2× cost savings without giving up quality.
Requests run with automatic fallback chains, so a provider outage never becomes your outage. ZDR-restricted requests only ever fall back to other verified ZDR endpoints — failover never becomes a compliance breach.
Every request appends an immutable audit event: the model used, the retention terms attested at that moment, the policies that fired, tokens, and cost. New memories are extracted with full provenance.
Platform
Multi-provider by design: Anthropic, OpenAI, Google Vertex, xAI, and open models via OpenRouter. Pin models by policy where it matters; let cost optimization choose everywhere else. The policy gate always decides what is eligible before economics decides what is best.
Org, department, and user namespaces with an explicit promotion workflow. PII never promotes upward. Departments cannot see each other’s memory. Every memory carries provenance — where it came from, when, and who may see it.
Route-time enforcement to verified ZDR endpoints, private mode on any individual request, and retention terms recorded per request. “Nothing was retained” becomes a fact you can prove, not a policy you hope held.
An append-only system of record for every AI interaction in your company: who asked, what ran, which policies fired, what it cost, and what the provider was contractually allowed to keep. The trail regulators and auditors demand.
Org and department budgets with hard caps enforced at the gate — not discovered on the invoice. Per-department usage is tracked on every request, so chargeback is a report, not a quarterly archaeology project.
Simple mode: a non-technical owner manages People, Privacy, Budget, and Templates in minutes. Advanced mode: hierarchy tree, policy editor, model registry, and audit center — built to govern thousands of seats.
Architecture
Pong is cloud-native on Google Cloud: a global anycast load balancer behind Cloud Armor, autoscaling Cloud Run services, and regional-HA PostgreSQL with point-in-time recovery. Tenant isolation is enforced with row-level security at the database — not just in application code.
The control plane and data plane are separate by design. On the enterprise tier, the data plane — the gateway and the memory store — runs in your own VPC: your prompts, memories, and audit events never leave your network. And with BYOK, model traffic runs on your provider keys and your contracts, never marked up.
Enterprise: control plane operated by Pong, data plane — gateway + memory — deployed inside your VPC.
Put a control plane between your company and the model providers — before the next budget, the next migration, or the next subpoena decides for you.