§01FAQ
Questions, answered plainly.
Everything enterprises ask before a Podar pilot — pricing, integration effort, quality, supported gateways, and how savings are proven.
General
- What is Podar?
- Podar is an enterprise AI cost-optimization platform. It sits between your applications and every major AI gateway, optimizes each prompt before execution, routes each request to the cheapest model that can answer it well, reuses semantically cached answers, and bills only a percentage of the spend it verifiably removes.
- How much can Podar save on our AI bill?
- In typical enterprise traffic mixes, 40 to 70 percent of AI spend is recoverable without any loss of quality. A conservative realized reduction of 30 percent on a $7M annual bill is about $2.1M, of which the customer keeps $1.575M after Podar's fee.
- How does Podar charge?
- A percentage of verified realized savings, typically 25 percent. There is no upfront license, no per-seat fee, and no platform minimum. If Podar saves nothing, Podar invoices nothing.
- What has to change in our applications?
- The base URL. Podar exposes a single drop-in compatible endpoint that speaks the same request and response format your SDKs already use.
- Will response quality drop?
- No. A quality bar is set per task class, every model swap is validated by an automated regression suite before it reaches production, and the Quality Index is reported next to savings in every review so cost cuts are never claimed in isolation.
- Which AI gateways and providers does Podar support?
- Podar orchestrates across OpenRouter, Portkey, TokenMix, Perplexity, Cloudflare AI Gateway, Kong AI Gateway, LiteLLM, Helicone, and Bifrost, plus direct provider connections, so customers never depend on a single supplier.
- How is Podar different from an AI gateway or an AI-FinOps tool?
- Gateways move traffic but do not treat spend reduction as an outcome. AI-FinOps tools report on spend but do not remove it. Hyperscaler dashboards only show their own consumption. Podar routes across all of them and is paid solely on the spend it deletes.
- How do you prove the savings you invoice?
- Every routed request is priced against the model that would otherwise have served it and logged with origin model, routed model, token counts, quality score, and realized saving. The invoice derives from that ledger and is auditable line by line.
- What is AI Yield?
- AI Yield is the share of AI spend that bought necessary work at an appropriately sized model. Its complement, AI Wastage, is the share spent overpaying for capability the task never needed or regenerating answers the organization already had. The two sum to 100 percent.
- What is the Podar Foundry?
- Podar Foundry is the second Podar engine. Where the Exchange routes each request to the best existing model, the Foundry watches recurring workloads, distills them into tiny task-specific Podar MicroModels, certifies them against a customer-defined quality bar, and adds them to the router as new, far cheaper destinations with automatic frontier fallback.
- Is Podar training its own frontier model?
- No. Podar distills tasks, not intelligence. It builds narrow specialists of a few hundred million to a few billion parameters for measurable task families such as extraction, classification, JSON transformation, PII detection, and short summarization — never a general-purpose mini model.
- How does a pilot start?
- With a free assessment. Podar reads two weeks of your AI traffic and returns your current AI Yield along with the dollar value of the waste that can be removed. There is no upfront license and no commitment to continue.
Model routing
- Does model routing reduce answer quality?
- No. Routing decisions are governed by a quality bar you set, and every model swap is validated by an automated regression suite before it reaches production traffic. If a cheaper model cannot hold quality on a task class, Podar does not route that class to it.
- Do we have to change our application code to use model routing?
- No. Podar exposes a single drop-in compatible endpoint. Applications keep the same request format; only the base URL changes.
- How much can model routing save on an enterprise AI bill?
- In typical enterprise traffic mixes, 40 to 70 percent of AI spend is recoverable without quality loss. A conservative realized reduction of 30 percent on a $7M annual bill is roughly $2.1M.
Prompt optimization
- What is prompt optimization?
- Prompt optimization reduces unnecessary tokens in AI requests before execution. It removes duplicated information, irrelevant conversation history, unused metadata, redundant examples, and overly verbose instructions while preserving user intent, required context, and constraints.
- Can prompt optimization break a carefully engineered prompt?
- Constraints, output schemas, system instructions, and user intent are protected. Optimization targets content that provably does not change the expected answer, and changes are validated against your regression suite.
Semantic caching
- How is semantic caching different from normal caching?
- Normal caching requires an exact match on the request. Semantic caching compares meaning, so 'what is our refund window' and 'how long do customers have to request a refund' can share one answer.
- How do you keep cached answers from going stale?
- Every cache class carries a time-to-live and can be invalidated on demand or by upstream data changes. Volatile categories can be excluded from caching entirely.
Adaptive workflow routing
- What is adaptive workflow routing?
- It is the decomposition of a complex AI request into specialized subtasks, routing each subtask to the most efficient capable model, validating each output, and recomposing the results into one final response.
- Does splitting a request add latency?
- Independent subtasks run in parallel, and most subtasks are handled by small fast models. In practice, end-to-end latency is typically comparable to or better than a single large-model call.
Multi-gateway fabric
- Which AI gateways does Podar support?
- OpenRouter, Portkey, TokenMix, Perplexity, Cloudflare AI Gateway, Kong AI Gateway, LiteLLM, Helicone, and Bifrost, with direct provider connections alongside them.
- Can we keep our existing gateway contract?
- Yes. Podar routes through the gateways and provider accounts you already have, and adds alternatives only where they are cheaper or more reliable.
AI Yield analytics
- What is AI Yield?
- AI Yield is the share of AI spend that bought necessary work at an appropriately sized model. Higher is better. Its complement, AI Wastage, is the share spent overpaying for unneeded capability or regenerating answers the organization already had.
- How is realized savings calculated?
- Each routed request is priced against the model that would otherwise have served it. The difference, minus the cost actually incurred, is the realized saving, and it is logged per request for audit.
Podar MicroModels
- What is a Podar MicroModel?
- A Podar MicroModel is a tiny, task-specific model fine-tuned from an open-weight base on one narrow, measurable workload — extraction, classification, JSON transformation, PII detection, short summarization. It becomes a new destination inside the Podar router, with confidence checks that escalate anything uncertain to a larger model.
- Is Podar building its own general-purpose model?
- No. Podar will not attempt to distill a general-purpose mini model. It builds a growing portfolio of narrow specialists that are certifiably good enough for one task family each.
- Which tasks are good distillation candidates?
- Excellent candidates: intent classification, entity extraction, prose-to-JSON, document categorization, PII redaction. Strong: standard support answers, policy matching, short domain summarization, text-to-SQL on a stable schema. Poor: open-ended strategy, complex multidisciplinary reasoning, novel legal or medical judgment, long-horizon agents, and current-events answers without retrieval.
- How do you handle the legal side of distillation?
- With a teacher-rights policy. Major providers' terms restrict using their outputs to train competing models, so Podar trains on open-weight models with permissive licenses, customer-owned labeled data, human-reviewed outcomes, negotiated distillation rights, or provider-hosted fine-tuning — and keeps data lineage for every training example.
- Could a MicroModel be retired?
- Yes. Podar stays economically neutral even toward its own models. If an external model becomes cheaper, or a task changes materially, traffic is reallocated or the MicroModel is retired.
Stop paying for AI overspend that buys you nothing.
A free assessment reads two weeks of your AI traffic and returns your current AI Yield and the dollar value of the waste we can remove.