If vLLM already solved LLM serving, why did SGLang appear?
Background
How ITQ scaled AI inference at enterprise volume
ITQ Technologies is the exclusive regional Travelport partner across India, Sri Lanka, the Maldives, and Bhutan, connecting thousands of agencies to global airline inventory. A core part of their business involves managing the ticket lifecycle, which requires the precise interpretation of fragmented airline change and cancellation policies.
To handle this, ITQ uses AI to automate the processing of these complex carrier rules.
As these services scaled to hundreds of billions of tokens per month, the primary obstacle became the infrastructure required to run inference reliably and economically.
Self-hosting open-weight models on general purpose cloud GPUs led to poor unit economics and unpredictable performance. Serverless frontier models were cost-prohibitive at ITQ’s token volumes. And shared, multi-tenant environments compromised both latency guarantees and intellectual property protection.
ITQ needed a fundamentally different approach to inference infrastructure.
The Opportunity
From general-purpose infrastructure to inference-optimized cloud
Airline change and cancellation rules are notoriously complex. Booking platforms must instantly decode thousands of policies across hundreds of carriers to display accurate charges. If a platform quotes a fee lower than the actual airline charge, the agency absorbs the difference.
To hedge against this, the industry defaults to quoting the highest possible fees, frustrating travelers and eroding sales.
ITQ saw an opportunity to replace this defensive pricing model with AI-driven precision: interpret fragmented airline policies in real time and produce commercially accurate pricing decisions at scale.
Realizing this vision required inference infrastructure that could deliver:
- Deterministic performance: Consistent, low-latency inference under production load to support real-time booking workflows.
- Sustainable token economics: A cost structure viable at hundreds of billions of tokens monthly, replacing the unsustainable economics of serverless frontier models.
- Sovereign infrastructure: A dedicated, private environment that protects proprietary models and fine-tuned logic as intellectual property.
- Solution partnership: An inference-optimized managed service with engineering support to bridge the gap between hardware and production-grade AI.
The Solution
How ITQ converts unstructured airline policies into actionable data
To deliver this precision, ITQ developed two specialized AI services that function as structured rule interpretation engines.
LLM Manual Fare Rule (MFR)
Specialized for interpreting complex airline policy documents and regulatory mandates like DGCA rules.
LLM VR3 (Refund Processing)
An automated engine for calculating exact refund and reissue figures across domestic and international scenarios.
The technical foundation of this suite is a custom Qwen 2.5 7B-Instruct model, specifically tuned to navigate the unique commercial logic and linguistic nuances of global airline filings.
For consistent performance under heavy loads, ITQ transitioned from shared public clouds to a dedicated, managed inference environment on Neysa Velocis.
The Neysa Partnership
Hardening the infrastructure for deterministic performance
ITQ partnered with Neysa to architect a secure, private inference environment on Neysa Velocis Cloud.
By hosting the Qwen 2.5 7B model on dedicated NVIDIA GPU clusters on Neysa’s cloud, ITQ achieved deterministic performance, with complete isolation from other tenants, protecting proprietary commercial logic and IP while delivering predictable unit economics.
Neysa’s engineering team hardened the system for production:
Speed and Simplicity
Spin up GPU-powered workloads in minutes. Reduce time-to-result with low-latency access and optimized pipelines that simplify everything from provisioning to inference. No multi-dashboard sprawl, no vendor lock-in.
Security and Control
Enterprise-grade, multi-layered security designed for regulated environments. Role-based access, policy enforcement, real-time audit trails, and zero-trust architecture, built in.
Scalability
From 1 to 1,000+ GPUs, scale as your models grow. Neysa Velocis is architected to support elastic scaling, hybrid burst, and workload-aware optimization.
Cost-Effectiveness
Stop burning cycles on idle infra. Transparent billing, spot-aware scheduling, and multi-tiered GPU pricing deliver superior tokens-per-dollar efficiency.
The Impact
Pioneering sovereign, cost-efficient AI for the travel industry
40% Reduction in Total Cost of Ownership vs. General Purpose cloud
Sustainable token economics through Neysa’s managed open-weights deployment.
Deterministic Quality of Service on Token Performance
Sub-2 second P99 latency for real-time fare rule and refund calculations within live booking flows.
141 Billion Tokens Processed Monthly
Production-scale AI sustained across agency and travel retail workflows.
Precision-tuned Qwen 2.5 7B-Instruct delivers high-fidelity results across complex airline policy variants, sourced from ITQ’s internal benchmarks.
Model IP Retained as Protected Work Product
Proprietary models and fine-tuned logic secured within a dedicated, private environment.
Take the next step
Ready to move your enterprise AI use cases to production? Contact our team today to request a technical demonstration and explore our managed inference solutions.
