Don't Use a Reasoning Model for a Simple Decision.
Zirnay is the high-speed decision layer for AI applications. Frontier LLMs excel at synthesis, deep research, and open prose. But routing a support ticket, scoring a credit card transaction, or verifying if an agent is authorized to take an action doesn't require generative reasoning. Zirnay handles discrete decisions in < 4.5ms uncached container P99 (down to < 2.5ms on fast-path route cache hits*) at up to ~97% lower cost than frontier generative models ($0.00010 vs $0.0030/call).
Every AI Request Doesn't Need a Frontier Model
Put a fast, deterministic decision layer between your application and your AI stack. Zirnay intercepts requests in milliseconds to score, gate, and route actions.
Refund card, flag fraud, trigger webhook without invoking any LLM.
Route routine query to 8B model or canned template (~90% cost savings).
Only invoke GPT-4o / Claude when multi-step synthesis is genuinely required.
AI Gateways Route Models. Zirnay Makes Decisions.
AI Gateways (OpenRouter, Portkey, RouteLLM) ask: "Which model should receive this prompt?"
Zirnay asks: "What business action should actually happen?"
Model Routing & Token Brokerage
Manages vendor endpoints, provider failovers, and streaming rate limits between OpenAI, Anthropic, and open models.
- "Is OpenAI currently returning 503?"
- "Which vendor has lower cost per 1M output tokens?"
- "Balance traffic 60/40 between Azure and Bedrock."
Semantic Decisions & Agent Guardrails
Evaluates incoming intent, computes calibrated risk probabilities, and decides whether tools, workflows, or reasoning models should execute.
- Gate: "Is this agent authorized to execute a $450 refund?" (YES / NO)
- Score: "What is the fraud probability of this credit swipe?" (0.94)
- Choice: "Which business department owns this ticket?" (Billing)
- Arbitrate: "Which candidate action wins the bidding auction?"
Pluggable System One Model Engines
Zirnay is the unified Decision Layer for AI. Underneath, task-specialized execution engines power each primitive:
Laya Decision Engine
Non-autoregressive ModernBERT 421M architecture with custom classification heads. Evaluates state and candidate labels directly into calibrated softmax distributions without token generation.
- Choice: Dynamic multi-option intent & classification
- Score: Continuous 0.0–1.0 risk and priority scoring
- Yes/No: High-throughput binary security gate
Strands Decider Routing Engine
Concurrent distributed decision-graph runtime for high-throughput edge dispatch, auction bidding, and multi-agent reflex mediation. Evaluates in 4.20ms via 3-plane container gateway (2.10ms cached), or sub-millisecond (0.85ms) when embedded on native edge nodes.
- Graph Triage: Weighted topological dispatch
- Dynamic Affinity: Multi-region routing policies
- Bandit Policies: Real-time exploration & arbitration
AnyJev Calibrated Engine
The breakthrough engine for zero-training decisions with empirical bias elimination. Turns open-weights foundation LLMs (integrated out-of-the-box with Qwen 2.5, Llama 3.1, and Mistral 7B) into deterministic, calibrated deciders by eliminating positional and label-order bias through cyclic permutation averaging and unconditional prior division under defined evaluation benchmarks.
- Cyclic Permutation: Eliminates position/order bias
- Prior Division: Eradication of label bias under cyclic testing
- Zero-Training (L0): Direct execution over Qwen, Llama, and Mistral weights
Which Engine Do You Need?
Use Laya when evaluating high-frequency semantic text, risk scores, or sub-5ms intent gates. Use Strands Decider for DAG routing graphs, auction bid triage, or distributed multi-agent mediation. Use AnyJev when requiring zero-training deployment over open-weights LLMs with guaranteed mathematical eradication of order and label bias.
When NOT to Use Zirnay
Zirnay is purpose-built for fast deterministic reflexes. Do not use Zirnay for open-ended creative prose generation, multi-page document summarization, code generation, or multi-turn conversational chat. Use frontier LLMs (Claude, GPT, Gemini) for those generative tasks.
How Zirnay Learns Your Domain
You can get started immediately with zero-shot evaluation, or upload labeled datasets to fine-tune specialized checkpoint heads:
Pass dynamic label options in your request payload (e.g. ["APPROVE", "CHALLENGE", "DECLINE"]). Laya projects input embeddings against label candidates in a single forward pass without model retraining.
Upload domain-specific labeled JSONL datasets via POST /v1/teams/:team/models/:model/train to fine-tune classification heads. Versioned checkpoints (chk_89f02c77) can be activated or rolled back with zero downtime.
Enterprise Data Privacy & Cache Architecture
We believe architectural transparency is non-negotiable for enterprise decision systems. Here is exactly how raw payloads, route caches, and telemetry are handled:
- Cache Key: One-way 16-hex SHA-256 digest:
route:{org}:{team}:{model}:{sha256}. Raw text is never in the key. - Cached Value: Stores only the categorical label & score (e.g.
APPROVE, 0.98), never the payload. - TTL: Ephemeral 300-second (5 min) volatile TTL in RAM.
- Can It Be Disabled? Yes. Send
Cache-Control: no-cacheorX-Bypass-Cache: trueto bypass cache entirely for 100% volatile RAM execution.
zrn_live_...), and all edge and intra-plane data in transit is protected with TLS 1.3 encryption.
Where the Decision Layer Wins in Production
Frontier reasoning models are indispensable for complex open-ended synthesis. But for high-velocity, mission-critical operations, enterprises require deterministic reflexes.
Financial Services AI Decision Infrastructure
Card authorization pipelines, instant payment routing, KYC/AML screening, and dispute triage process thousands of queries per second. They cannot tolerate 400ms+ LLM latency or expose PII to external log warehouses.
- Sub-5ms Deterministic Reflex: Instant APPROVE / CHALLENGE / DECLINE gates within card network SLAs.
- Zero Raw Payload Retention: Card numbers and transaction state evaluate solely in volatile RAM.
- Platt-Calibrated Risk (ECE < 0.03): Mathematically reliable confidence scoring for automated risk thresholds.
- Deterministic Audit Replay: Cryptographically verified decision telemetry for compliance audits.
AI Agent Guardrails & Runtime Decision Layer
Autonomous agentic loops (LangChain, CrewAI, AutoGen) stall and bleed capital when invoking $0.003 frontier reasoning models for basic tool gating, loop validation, or intent routing at every single step.
- Sub-5ms Tool Authorization: Authorize refunds, database mutations, or API calls before execution.
- Candidate Action Arbitration: Strands Decider evaluates candidate tool paths in a single graph pass.
- Native MCP Integration: Direct Model Context Protocol tool endpoints for agent runtime integration.
- 30x Cost Reduction: Strip out token generation overhead from non-generative validation steps.
Full-Stack Decision Plane Architecture
Turning raw model weights into enterprise-grade decision systems requires more than isolated container scripts. Zirnay provides a 12-component Decision Plane out of the box:
No $2.50/hr GPU reservations or idle waste. Delivers 4.20ms P99 on commodity CPU pods with instant horizontal auto-scaling.
One-way 16-hex SHA-256 keyed Redis routing cache delivering 2.5ms fast-path responses without storing customer text.
Raw logits are temperature Platt-scaled to guaranteed true probabilities (ECE < 0.03) for strict automated policy gates.
Supply dynamic candidate label arrays in runtime payloads without model retraining or weight updates.
Eliminates position and label bias across open-weights LLMs via cyclic permutation averaging and unconditional prior division.
Supervised fine-tuning pipeline with versioned checkpoints (chk_...) hot-swapped or reverted instantly.
Segregated Control Plane (RBAC/quotas), Routing Plane (triage/cache), and Inference Plane (low-jitter execution).
Strict RAM-only execution. Inputs are never saved to disk, relational databases, or training datasets.
First-class Model Context Protocol (MCP) tool server, Python SDK, and TypeScript client for seamless agent integration.
Live Prometheus telemetry tracks entropy shifts, confidence distribution drift, and edge anomalies automatically.
Cryptographic isolation across organizations and teams with scoped API keys (zrn_live_...) and TLS 1.3 encryption.
Triage incoming queries seamlessly across deterministic code, small fast models, and frontier reasoning LLMs.
Interactive Cost & ROI Calculator
Compare baseline frontier models against Zirnay engines: Laya, Strands Decider, and AnyJev.
*First 100,000 decisions/mo are free across all engines. Click any engine above to inspect its 3-plane latency and economics.
3-Plane Architecture: Single forward-pass non-autoregressive decision head across Control, Data, and Inference planes. Uncached container CPU executes in 4.20ms P99 (realistic for unique payloads); fast-path route cache responds in 2.10ms P99 (ephemeral 300s hashed Redis cache, bypassable via Cache-Control: no-cache).
Tested Latency Curve & 3-Plane Architecture
We believe in transparent engineering math. In our disaggregated 3-plane architecture (Control, Data Gateway, Inference Pod), server execution ranges from 1.84ms on dedicated GPU to ~4.2ms on container CPU, with < 15ms regional edge transit. Below is our measured benchmark matrix.
| Payload Size | Representative Use Case | NVIDIA L4 (TensorRT FP16) | Container CPU (3-Plane Gateway) | Regional Network SLA |
|---|---|---|---|---|
| Short Reflex (≤ 64 tokens) | Card auth flag, KYC binary gate, agent intent triage | 1.84 ms P99 | 4.20 ms P99 (2.10 ms cached) | < 15.0 ms P99 SLA |
| Medium Context (256 tokens) | Support email triage, incident alert log scoring | 3.80 ms P99 | 6.30 ms P99 | < 20.0 ms P99 SLA |
| Full Document (1,024 tokens) | Multi-paragraph customer claim or contract clause | 11.20 ms P99 | 28.40 ms P99 | < 35.0 ms P99 SLA |
*Methodology: Measured across disaggregated 3-plane architecture: Control Plane (auth & policy), Data Plane Gateway (Express + Redis route cache), and Python Inference Pod (non-autoregressive ModernBERT). Uncached latency reflects novel unique inputs; route cache hits use ephemeral 300s tenant-scoped SHA-256 digests in Redis (bypassable via Cache-Control: no-cache). GPU numbers reflect dedicated L4 TensorRT acceleration; CPU reflects standard container deployment.
Preliminary Quality Benchmark: Banking Intent Triage
Task: Fine-grained banking intent triage across 77 domain classes (Banking77 Benchmark, 3,080 test queries, 128-token avg length).
| Evaluation Metric | Zirnay Laya (ModernBERT Head) | Frontier LLM (5-Shot Prompted) | Tuned GBDT (XGBoost + Embeddings) |
|---|---|---|---|
| Top-1 Intent Accuracy | 93.8% | 92.4% | 88.6% |
| P99 Server Latency | 1.84 ms (TensorRT L4) / 4.20 ms (3-Plane CPU) | 480 ms P50 / 680 ms P99 | 18.50 ms (c3-standard-4) |
| Cost Basis (per 1M decisions) | $100.00 (Growth Tier) | $3,000.00 (Token API pricing) | ~$240.00 (Dedicated VM infra) |
| Zero-Shot Label Candidates | Supported (Dynamic options vector) | Supported (High prompt tokens) | Unsupported (Retraining required) |
| Raw Payload Retention | Zero (Volatile RAM execution; ephemeral 300s SHA-256 route cache bypassable via Cache-Control: no-cache) | Subject to vendor log policy | Stored in logging warehouse |
*Reproducibility: Evaluated on test split of Banking77 benchmark. Frontier LLM evaluated using 5-shot prompt on GPT-4o mini streaming endpoint. GBDT baseline tuned with Optuna over frozen dense embeddings (dense vector extraction + tree traversal).
Ready for Instinctual AI at Production Scale?
Get started with 100,000 free decision events every month. Connect via standard REST, Model Context Protocol (MCP), or Python/TypeScript SDKs.