CARDIAC-PURR AI Control Plane

CARDIAC-PURR AI Control Plane sits between your app and LLM providers as a drop-in proxy: swap the OpenAI SDK's base_url and it routes each request to the cheapest model tier (small/medium/large) that meets a pre-set quality bar, across Anthropic, OpenAI, Google, Azure OpenAI, Mistral, Cohere, DeepSeek, Grok, Llama and custom endpoints. Around that routing core it adds policy checks, confidence-threshold blocking, automatic failover, and a per-request audit trail — positioned as a governance layer, not just a router. Source: ^[4]

Free tier: $0 forever, up to 15,000 requests/month, then routing pauses (HTTP 429) until the next cycle. Growth: 25% of verified savings, capped at $75,000/month. Scale: 22% of verified savings, capped at $150,000/month. Enterprise: negotiated. There's no flat subscription tier at all — if the router saves you nothing, you pay nothing, and the FAQ works through billing edge cases in real detail: retried/failed-over calls are billed once against whichever provider produced the successful response, fee caps stop the fee but not the routing, and self-serve cancellation gives 30 days to export data before anonymisation. That level of specificity in a public FAQ is unusual for a vendor this size. The trade-off: 'verified savings' is defined and measured by CARDIAC-PURR itself via its own 'Model Delta Invoice' — there's no independent audit of that number, so a buyer is trusting the vendor's math on what they're being billed a percentage of. Source: ^[1]

The published benchmark (100 real queries per provider across 9 providers, 4 verticals, run July 12, 2026) reports routing accuracy of 98-100% and cost savings ranging from 88.3% (DeepSeek) down to 30.9% (Grok). The low Grok number is left in and explained rather than dropped — Grok's three tiers currently call the same underlying model, so there's no cheap-tier saving to route to, which the vendor states plainly rather than hiding behind an average. The methodology page also discloses Wilson-score confidence intervals, that Cohere's 98% (not 100%) accuracy reflects two provider-side timeouts rather than routing error, and that the benchmark is self-run, not third-party audited. A separate August 12, 2026 agentic-framework benchmark (350 requests across LangChain, LlamaIndex, AutoGen, CrewAI, Haystack, DSPy, Semantic Kernel) reports 100% accuracy / 80% savings on a single factual-QA workload at small-tier routing only — the vendor is explicit this doesn't generalize to other workloads or tiers. Sources: ^[3]

The vendor's own compliance page states plainly that neither SOC 2 Type I (target Q1 2027) nor ISO 27001 (target Q4 2026) is complete, that at-rest encryption specifics aren't yet documented, and that no third-party penetration testing has been done. What is in place: TLS 1.2+ in transit, role-based access control, an EU AI Act Article 6 self-assessment (dated 25 June 2026) concluding the product isn't an Annex III high-risk system, primary hosting in Frankfurt with EU SCCs for any data leaving the EEA, and a 99.5% uptime target whose remedy for a breach is contract termination with 30 days' notice rather than service credits (credits require a separately negotiated Schedule C). No HIPAA BAA is offered — the vendor states this was a deliberate removal. Source: ^[2]

2info Sp. z o.o., Krakow, Poland — a two-person team per the vendor's own about page: founder Emil Igidov (described as a business/technology lawyer practicing across Sweden, Denmark, Poland) and co-founder Abdullah Ok (a test automation engineer at Axis Communications, handling backend). The company's GitHub org was created 29 May 2026; its one public repo is 3KB, has 0 stars, and hasn't been touched since it was created on 2 August 2026 — it's a placeholder, not an open-source component of the product. Two USPTO provisional patent filings are claimed (GPU thermal family, and a separate AI-execution-platform family filed May 2026) — provisional applications aren't published by the USPTO, so this can't be independently verified beyond the vendor's own statement. An extensive search turned up no independent press coverage, reviews, or third-party mentions of CARDIAC-PURR, 2info, or Emil Igidov outside the vendor's own site; the accolades listed (Constructor Start 2026 semi-finalist, Polish Startup Hunt 2026, OVH Cloud grant, Dubai Create Apps Championship recognition) are startup-program selections, not customer or revenue evidence. Sources: ^[5], ^[7], ^[6]