Longinus Research.

Choose a Product Deep Dive

Three specialized engines, built for production SLAs

Click any product below for detailed enterprise use cases, architecture, integration snippets, and deployment paths.

01 · CoreRank

Enterprise Reranking & Adversarial-Safe Judging

Eliminate RAG ranking noise, keyword-stuffing exploits, and LLM-as-a-Judge position bias in one drop-in API call.

  • #1 overall on 5-BEIR (0.4326 NDCG@10)
  • 90.3% adversarial enterprise accuracy
  • 1.8% flip rate & 0.0% position bias
  • Cohere /v1/rerank, OpenRouter & Private VPC
Open CoreRank Product Page →
02 · CoreCompile

Sub-Millisecond Agent Guardrails & Tool Routing

Drop-in OpenAI-compatible proxy that offloads recurring prompt-injection firewalls, hallucination gates, and multi-tool routing.

  • 0.32 ms CPU firewall (99.4% AUC, 742×)
  • 100.0% BFCL Top-1 tool routing in 8.88 ms
  • 85.5% HaluBench groundedness gate in 4.50 ms
  • Automatic upstream fallback on uncertain inputs
Open CoreCompile Product Page →
03 · CoreKV

4×–8× Long-Context KV Memory & Serving Engine

Break through GPU memory walls for long-context chat, code assistants, and multi-document agents without buying more GPUs.

  • 4×–8× KV cache memory reduction (3.58 GB → 448 MB)
  • Up to 3.07× faster decode attention vs. PyTorch SDPA
  • 2.03× max-batch throughput (24 streams @ 4K ctx)
  • 100% 1K–8K NIAH & Multi-Needle · 95% CoT Math at 4×
Open CoreKV Product Page →

Buyer’s Guide

Match your production bottleneck to the right engine

Every product is deployed standalone in days—no forced platform lock-in.

Primary Pain Point Product Headline Metric Deployment Path
RAG retrieves wrong passages or gets gamed by boilerplate CoreRank #1 5-BEIR · 1.8% flips · 0.0% bias Managed API / Cohere SDK / VPC Container
Agent guardrails & routing add 150–400 ms and double API bills CoreCompile 0.32 ms CPU (742×) · 100% Top-1 routing 2-line OpenAI Proxy / Kubernetes Sidecar
Long-context chat & code serving hits GPU OOM limits CoreKV 4×–8× KV cut · 3.07× attn · 2.03× batch Self-contained Serving Runtime / Custom Conversion

Full-Stack Production Blueprint

How all three products combine on a single enterprise query

Start with one product, then connect the rest to eliminate bottlenecks from ingress security to long-context generation and egress verification.

Step 1 · Ingress Firewall & Routing

CoreCompile Edge Proxy

Blocks prompt injections in 0.30 ms on CPU and selects the right tool or retrieval index in 5.11 ms before calling any heavy model.

Step 2 · Adversarial-Safe Retrieval

CoreRank Reranker

Re-orders candidate RAG documents with 0.0% position bias and 1.8% flip rate so only clean, relevant context reaches generation.

Step 3 · High-Concurrency Generation

CoreKV Serving Engine

Generates long-context answers in 4×–8× less KV memory, packing up to 3× more concurrent sessions per GPU.

Step 4 · Egress Groundedness Gate

CoreCompile Verification

Checks every response for unsupported claims in 2.03 ms, triggering automatic upstream fallback or retry only when confidence drops.

Zero-Risk Rollout

From shadow benchmark to production VPC in 3 weeks

Week 1 · Shadow Eval

Measure on Live Traffic

Run non-blocking shadow scoring on your real queries, agent routes, and long-context logs to validate exact ROI before any production cutover.

Week 2 · Private VPC Cutover

Drop-In Container & SDK

Deploy hardened Docker / Kubernetes containers inside your AWS, GCP, Azure, or on-prem cluster with zero external data egress.

Week 3 · Enterprise Scale

SLA & Ongoing Optimization

Lock in confidence thresholds, automatic fallback rules, and enterprise SLA monitoring across your production fleet.

Direct Pilot Request

Start a 1-week shadow benchmark on your real workload

Tell us your primary bottleneck and deployment target (Cloud API or Private VPC). Zero data leaves your infrastructure during VPC pilots.

Or email directly: marcio.diaz@gmail.com