We migrated TrueStay to AWS Bedrock, cut their AI spend by 90%, and contributed to 4x user growth

Our team at allOps recognized the opportunity to include TrueStay in a top-tier AWS for Startups program that offsets the costs for growing ventures while giving them the stability of the giants.

AWS Bedrock gave TrueStay access to top-tier Claude models, platform-level security, and predictable, optimized costs. Render, unfortunately, couldn’t provide this.

We carried out a complete migration to AWS, with a focus on AI workloads:

• We secured access to the AWS for Startups program with significant credits, which drastically reduced the direct costs of AI calls and covered them with credits.

• We took over the entire infrastructure, hosting, and operations from Render and moved everything to AWS (ECS containers, EC2, RDS database, Elastic Load Balancing), with full control over scaling, regions, and costs.

• We moved the complete agent reasoning and review analysis to Claude models via AWS Bedrock. Critically, this brought AI inference inside the AWS service boundary, where it is now governed by the same IAM policies, VPC controls, CloudTrail audit trails, and cost allocation tooling as every other part of the infrastructure.
  
  Authentication moved to a secure, centralized model (IAM), eliminating the fragmented access management of the previous setup.

• We enabled Bedrock Guardrails as a platform-level safeguard (content filtering, PII redaction, prompt injection and attack detection, denied topic policies, and hallucination detection via Automated Reasoning checks)

None of these controls were available through direct external provider endpoints. For a travel platform handling user-generated queries and third-party review content, this is a material gain in product quality and compliance.

But credits are only part of the story.

Across TrueStay’s architecture,has become a core efficiency layer, operating at multiple levels of the stack.

At the AI layer, we implemented Bedrock Prompt Caching to store frequently reused context, long system prompts, agent reasoning scaffolding, across calls, cutting costs by up to 90% and latency by up to 85%.

Given TrueStay’s workload of repeated review analysis over shared prompt context, this compounds with scale.

Beyond prompt-level caching, we introduced Amazon ElastiCache to persist the results of previous search runs and analytical computations. When a user queries a property category, market segment, or date range that was recently processed, the system serves the cached result directly, with no redundant model invocations or repeated data retrieval.

This is especially impactful for high-frequency lookups and dashboard refreshes where the underlying data hasn’t changed.

We also moved model selection from a static swap to a dynamic layer. Rather than simply replacing Haiku with Sonnet, Bedrock Intelligent Prompt Routing automatically directs simpler queries to lighter, more cost-effective models and routes complex multi-step reasoning to Sonnet 4.6, requiring no application-level logic, and delivering up to 30% lower cost without compromising response quality.

Automated prompt optimization to eliminate manual work

Finally, we put continuous improvement on rails with Bedrock Advanced Prompt Optimization. This feature benchmarks and automatically rewrites prompts across up to five models at once using sample inputs, expected outputs, and defined evaluation metrics.

It returns evaluation scores, cost estimates, and latency comparisons against the originals.

For TrueStay’s reasoning and review workloads, this replaces manual prompt-engineering iteration and keeps prompt quality in step with model updates and product growth, without dedicated engineering effort each cycle.