Projected 3-year AI/ML savings with self-hosted open-source models
Cut from the migration timeline by provisioning waves in parallel
Agent rewrites, with Bedrock reached through a one-line config change
Lovelace
Pittsburgh, Pennsylvania
Financial Services / Data Intelligence
Amazon EKS, Amazon Aurora PostgreSQL, Amazon S3, AWS Bedrock, Amazon SageMaker, Amazon ElastiCache, Amazon Redshift, AWS WAF, AWS Secrets Manager, AWS KMS, Self-hosted observability (Grafana, Mimir, Loki, Tempo)
Lovelace, a data intelligence company operating AI-powered platforms for financial services and multi-tenant enterprise customers, engaged Avahi to conduct a comprehensive assessment for migrating their entire infrastructure and AI/ML workloads from Google Cloud Platform to AWS. The assessment revealed that Lovelace’s Kubernetes-native architecture was already substantially cloud-agnostic, with the real migration complexity concentrated in the AI/ML layer (Vertex AI models, fine-tuned inference, and embedding pipelines) and the data tier (Cloud SQL, Bigtable, and BigQuery). Avahi delivered a complete migration blueprint covering target architecture, model benchmarking, cost projections, and a phased wave plan that reduces the migration timeline which is about 14 to 22 weeks through parallel provisioning.
Lovelace operates two primary platforms: a data intelligence platform that processes news and financial data through 35 external feeds, and a multi-tenant platform serving enterprise customers. Their platform runs on self-hosted, in-cluster databases, with heavy reliance on Google Vertex AI for LLM inference, embeddings, and fine-tuned models. The platforms support high-throughput workloads, with some components requiring up to 2,000 queries per second for entity resolution and embedding search operations.
Lovelace needed to migrate their entire cloud infrastructure from GCP to AWS, but the complexity of their AI/ML workloads created significant uncertainty around cost, timeline, and technical feasibility. Their platforms depended on Vertex AI for multiple model types (Gemini variants for agent orchestration, text-embedding-005 for vector search, and custom fine-tuned models for entity resolution), and the migration path for each was unclear.
The fine-tuned models presented particular challenges: they were critical to the entity resolution pipeline but had no direct AWS equivalent, requiring evaluation of Amazon Nova, SageMaker, and self-hosted options. Without a clear understanding of the migration path, Lovelace faced the risk of extended dual-cloud costs, potential service disruptions during cutover, and the possibility of discovering blocking technical issues mid-migration.
Additionally, the cost implications of switching AI/ML providers were unknown. Initial assumptions suggested that moving to AWS Bedrock would be straightforward, but without detailed analysis of token volumes, QPS requirements, and model pricing differences, Lovelace could not make informed decisions about their target architecture.
Lovelace selected AWS as their target cloud platform because they want to get in the AWS Marketplace ecosystem. Additionally, they can take advantage of AWS’s breadth of AI/ML services. AWS Bedrock offered access to multiple foundation models through a unified API, while Amazon SageMaker provided a managed path for their fine-tuned model training and inference requirements. The combination of Amazon EKS for their Kubernetes workloads, Aurora PostgreSQL for their relational data tier, and the AWS AI/ML stack created a cohesive target architecture that could support their growth while reducing operational complexity.
Lovelace required a partner with deep expertise in both AI/ML systems and cloud migration to navigate the complexity of their multi-platform environment. Avahi’s experience with AWS MAP (Migration Acceleration Program) engagements provided the structured methodology needed to assess, plan, and execute a migration of this scale. Avahi’s ability to conduct code-level analysis of the existing architecture, benchmark candidate models against production workloads, and deliver actionable cost projections gave Lovelace the confidence to proceed with a migration that touched every layer of their technology stack.
Avahi conducted a comprehensive assessment spanning AI/ML workloads, infrastructure components, and data tier services across all of Lovelace’s GCP projects.
AI/ML Architecture Analysis
The assessment began with a detailed inventory of every model, embedding provider, and AI service across both platforms. Avahi’s engineers performed code-level validation to map each component to its actual model dependencies, correcting several assumptions from earlier discovery passes. A key finding was that Lovelace’s agent framework (built on Google ADK) was model-agnostic by construction: ADK ships a LiteLLM adapter that lets every Python-based agent reach AWS Bedrock with a single-line configuration change, eliminating the need for agent rewrites.
For the Go-based services that consumed the majority of token volume (Resolve Matcher, Resolve Searcher, Agents Serve, and Fetch Pipeline), Avahi designed a native Bedrock client implementation following the existing provider pattern in the codebase. This approach maintained the existing abstraction layer while adding Bedrock as a new backend option.
Model Benchmarking and Selection
Avahi developed a purpose-built benchmarking framework to evaluate candidate models on the two NLP tasks most critical to Lovelace’s platform: entity extraction and entity resolution. The framework captured quality metrics (precision, recall, F1), performance metrics (latency percentiles, time-to-first-token, requests per second), and cost per million tokens for each candidate.
The benchmarking revealed that Amazon Nova 2.0 Lite, while matching Gemini quality on some workloads, would more than double total AI/ML costs over three years due to pricing differences on high-volume components. Open-source self-hosted models (Gemma 4 31B and gpt-oss-120b) emerged as the only path that reduced costs relative to the current GCP baseline, with projected three-year savings of approximately $240,000.
Infrastructure Assessment
The infrastructure assessment confirmed that Lovelace’s “everything is Kubernetes” architecture was already substantially cloud-agnostic. Zero Cloud Run, zero GCE VMs, and zero Dataflow services were found across either platform. Databases ran as self-hosted, in-cluster deployments rather than managed GCP services, meaning they could migrate with the same deployment configuration to EKS.
The real GCP dependency surface reduced to: Vertex AI (covered by the AI/ML assessment), GCS (direct migration to S3), Cloud SQL PostgreSQL (replatform to Aurora), Bigtable (replatform to ElastiCache, as code analysis confirmed it served only as an L2 cache), and BigQuery (replatform to Redshift or Athena, pending query pattern analysis).
Data Tier Recommendations
Avahi recommended Aurora PostgreSQL over RDS for the relational data tier, as Aurora Global Database was required for the cross-region DR strategy. For the self-hosted, high-throughput data store, Avahi designed a self-hosted deployment on EKS using dedicated, tainted node pools with local-NVMe instances (i3en or i4i), avoiding Karpenter’s consolidation logic that could terminate nodes and lose data. The vector store was confirmed to migrate with zero code change using the same deployment configuration on EKS.
Migration Wave Plan
Avahi developed a four-wave migration plan with specific technical gates and stability criteria for each wave transition. The plan estimated the total migration timeline is about 14 to 22 weeks through parallel infrastructure provisioning and parallel tenant cutovers in Wave 4. Each wave transition required 72-hour stability windows (availability above 99.5%, no incidents), with Wave 2 additionally requiring GCS-to-S3 sync with zero egress for seven consecutive days.
Cost Modeling
The assessment delivered two compute sizing bases: a ceiling estimate ($126,637 per month) sized to GCP’s autoscaler maximums to protect the funding case, and an active-usage floor ($38,496 per month) based on live GCP node counts. The active count of 65 nodes independently matched a separate LucidScale discovery figure from months earlier, validating the methodology. Total Year 1 infrastructure costs were projected at $1,890,090 at maximum utilization.
Complete AI/ML assessment report covering model inventory, migration strategies, and effort estimates for all components across both platforms
Infrastructure assessment report with target architecture, component matrix, and data tier recommendations
Model benchmarking framework and results for entity extraction and resolution tasks across candidate models
Three-year AI/ML cost projection comparing Nova 2.0 Lite, open-source self-hosted, and current GCP baseline
Four-wave migration plan with technical gates, stability criteria, and rollback procedures
Bedrock quota sizing analysis with specific quota increase requests required before cutover
Self-hosted data store deployment specification including instance types, node counts, and taint/affinity configuration
DR strategy cost matrix with RTO/RPO targets and monthly cost deltas for each option
The assessment provided Lovelace with a complete, validated migration blueprint that de-risked their GCP-to-AWS transition. The code-level analysis corrected multiple assumptions from earlier discovery passes, including the discovery that the first-pass cross-encoder model (previously estimated at 48GB VRAM) actually runs on CPU in production, eliminating a potential GPU infrastructure requirement.
The wave plan’s parallel provisioning approach reduced the projected migration timeline by 12 to 16 weeks compared to a sequential approach, minimizing the period of dual-cloud costs and operational complexity.
Projected three-year AI/ML cost with open-source self-hosted models: $1,367,747 (14.9% reduction vs. GCP baseline of $1,607,781)
Projected three-year AI/ML cost with Nova 2.0 Lite: $3,255,506 (102.5% increase vs. GCP baseline)
Migration timeline with parallel provisioning: 14 to 22 weeks (vs. 26 to 38 weeks sequential)
Active compute utilization: 65 nodes (33% of 197-node autoscaler ceiling)
Vector store capacity: approximately 200 million vectors across 12 shards with replication factor of 2
Peak QPS requirements validated: 2,000 QPS for embedding search, 500 QPS for agent inference
Lovelace
Pittsburgh, Pennsylvania
Financial Services / Data Intelligence
Amazon EKS, Amazon Aurora PostgreSQL, Amazon S3, AWS Bedrock, Amazon SageMaker, Amazon ElastiCache, Amazon Redshift, AWS WAF, AWS Secrets Manager, AWS KMS, Self-hosted observability (Grafana, Mimir, Loki, Tempo)
Lovelace, a data intelligence company operating AI-powered platforms for financial services and multi-tenant enterprise customers, engaged Avahi to conduct a comprehensive assessment for migrating their entire infrastructure and AI/ML workloads from Google Cloud Platform to AWS. The assessment revealed that Lovelace’s Kubernetes-native architecture was already substantially cloud-agnostic, with the real migration complexity concentrated in the AI/ML layer (Vertex AI models, fine-tuned inference, and embedding pipelines) and the data tier (Cloud SQL, Bigtable, and BigQuery). Avahi delivered a complete migration blueprint covering target architecture, model benchmarking, cost projections, and a phased wave plan that reduces the migration timeline which is about 14 to 22 weeks through parallel provisioning.
Lovelace operates two primary platforms: a data intelligence platform that processes news and financial data through 35 external feeds, and a multi-tenant platform serving enterprise customers. Their platform runs on self-hosted, in-cluster databases, with heavy reliance on Google Vertex AI for LLM inference, embeddings, and fine-tuned models. The platforms support high-throughput workloads, with some components requiring up to 2,000 queries per second for entity resolution and embedding search operations.
Lovelace needed to migrate their entire cloud infrastructure from GCP to AWS, but the complexity of their AI/ML workloads created significant uncertainty around cost, timeline, and technical feasibility. Their platforms depended on Vertex AI for multiple model types (Gemini variants for agent orchestration, text-embedding-005 for vector search, and custom fine-tuned models for entity resolution), and the migration path for each was unclear.
The fine-tuned models presented particular challenges: they were critical to the entity resolution pipeline but had no direct AWS equivalent, requiring evaluation of Amazon Nova, SageMaker, and self-hosted options. Without a clear understanding of the migration path, Lovelace faced the risk of extended dual-cloud costs, potential service disruptions during cutover, and the possibility of discovering blocking technical issues mid-migration.
Additionally, the cost implications of switching AI/ML providers were unknown. Initial assumptions suggested that moving to AWS Bedrock would be straightforward, but without detailed analysis of token volumes, QPS requirements, and model pricing differences, Lovelace could not make informed decisions about their target architecture.
Lovelace selected AWS as their target cloud platform because they want to get in the AWS Marketplace ecosystem. Additionally, they can take advantage of AWS’s breadth of AI/ML services. AWS Bedrock offered access to multiple foundation models through a unified API, while Amazon SageMaker provided a managed path for their fine-tuned model training and inference requirements. The combination of Amazon EKS for their Kubernetes workloads, Aurora PostgreSQL for their relational data tier, and the AWS AI/ML stack created a cohesive target architecture that could support their growth while reducing operational complexity.
Lovelace required a partner with deep expertise in both AI/ML systems and cloud migration to navigate the complexity of their multi-platform environment. Avahi’s experience with AWS MAP (Migration Acceleration Program) engagements provided the structured methodology needed to assess, plan, and execute a migration of this scale. Avahi’s ability to conduct code-level analysis of the existing architecture, benchmark candidate models against production workloads, and deliver actionable cost projections gave Lovelace the confidence to proceed with a migration that touched every layer of their technology stack.
Avahi conducted a comprehensive assessment spanning AI/ML workloads, infrastructure components, and data tier services across all of Lovelace’s GCP projects.
AI/ML Architecture Analysis
The assessment began with a detailed inventory of every model, embedding provider, and AI service across both platforms. Avahi’s engineers performed code-level validation to map each component to its actual model dependencies, correcting several assumptions from earlier discovery passes. A key finding was that Lovelace’s agent framework (built on Google ADK) was model-agnostic by construction: ADK ships a LiteLLM adapter that lets every Python-based agent reach AWS Bedrock with a single-line configuration change, eliminating the need for agent rewrites.
For the Go-based services that consumed the majority of token volume (Resolve Matcher, Resolve Searcher, Agents Serve, and Fetch Pipeline), Avahi designed a native Bedrock client implementation following the existing provider pattern in the codebase. This approach maintained the existing abstraction layer while adding Bedrock as a new backend option.
Model Benchmarking and Selection
Avahi developed a purpose-built benchmarking framework to evaluate candidate models on the two NLP tasks most critical to Lovelace’s platform: entity extraction and entity resolution. The framework captured quality metrics (precision, recall, F1), performance metrics (latency percentiles, time-to-first-token, requests per second), and cost per million tokens for each candidate.
The benchmarking revealed that Amazon Nova 2.0 Lite, while matching Gemini quality on some workloads, would more than double total AI/ML costs over three years due to pricing differences on high-volume components. Open-source self-hosted models (Gemma 4 31B and gpt-oss-120b) emerged as the only path that reduced costs relative to the current GCP baseline, with projected three-year savings of approximately $240,000.
Infrastructure Assessment
The infrastructure assessment confirmed that Lovelace’s “everything is Kubernetes” architecture was already substantially cloud-agnostic. Zero Cloud Run, zero GCE VMs, and zero Dataflow services were found across either platform. Databases ran as self-hosted, in-cluster deployments rather than managed GCP services, meaning they could migrate with the same deployment configuration to EKS.
The real GCP dependency surface reduced to: Vertex AI (covered by the AI/ML assessment), GCS (direct migration to S3), Cloud SQL PostgreSQL (replatform to Aurora), Bigtable (replatform to ElastiCache, as code analysis confirmed it served only as an L2 cache), and BigQuery (replatform to Redshift or Athena, pending query pattern analysis).
Data Tier Recommendations
Avahi recommended Aurora PostgreSQL over RDS for the relational data tier, as Aurora Global Database was required for the cross-region DR strategy. For the self-hosted, high-throughput data store, Avahi designed a self-hosted deployment on EKS using dedicated, tainted node pools with local-NVMe instances (i3en or i4i), avoiding Karpenter’s consolidation logic that could terminate nodes and lose data. The vector store was confirmed to migrate with zero code change using the same deployment configuration on EKS.
Migration Wave Plan
Avahi developed a four-wave migration plan with specific technical gates and stability criteria for each wave transition. The plan estimated the total migration timeline is about 14 to 22 weeks through parallel infrastructure provisioning and parallel tenant cutovers in Wave 4. Each wave transition required 72-hour stability windows (availability above 99.5%, no incidents), with Wave 2 additionally requiring GCS-to-S3 sync with zero egress for seven consecutive days.
Cost Modeling
The assessment delivered two compute sizing bases: a ceiling estimate ($126,637 per month) sized to GCP’s autoscaler maximums to protect the funding case, and an active-usage floor ($38,496 per month) based on live GCP node counts. The active count of 65 nodes independently matched a separate LucidScale discovery figure from months earlier, validating the methodology. Total Year 1 infrastructure costs were projected at $1,890,090 at maximum utilization.
Complete AI/ML assessment report covering model inventory, migration strategies, and effort estimates for all components across both platforms
Infrastructure assessment report with target architecture, component matrix, and data tier recommendations
Model benchmarking framework and results for entity extraction and resolution tasks across candidate models
Three-year AI/ML cost projection comparing Nova 2.0 Lite, open-source self-hosted, and current GCP baseline
Four-wave migration plan with technical gates, stability criteria, and rollback procedures
Bedrock quota sizing analysis with specific quota increase requests required before cutover
Self-hosted data store deployment specification including instance types, node counts, and taint/affinity configuration
DR strategy cost matrix with RTO/RPO targets and monthly cost deltas for each option
The assessment provided Lovelace with a complete, validated migration blueprint that de-risked their GCP-to-AWS transition. The code-level analysis corrected multiple assumptions from earlier discovery passes, including the discovery that the first-pass cross-encoder model (previously estimated at 48GB VRAM) actually runs on CPU in production, eliminating a potential GPU infrastructure requirement.
The wave plan’s parallel provisioning approach reduced the projected migration timeline by 12 to 16 weeks compared to a sequential approach, minimizing the period of dual-cloud costs and operational complexity.
Projected three-year AI/ML cost with open-source self-hosted models: $1,367,747 (14.9% reduction vs. GCP baseline of $1,607,781)
Projected three-year AI/ML cost with Nova 2.0 Lite: $3,255,506 (102.5% increase vs. GCP baseline)
Migration timeline with parallel provisioning: 14 to 22 weeks (vs. 26 to 38 weeks sequential)
Active compute utilization: 65 nodes (33% of 197-node autoscaler ceiling)
Vector store capacity: approximately 200 million vectors across 12 shards with replication factor of 2
Peak QPS requirements validated: 2,000 QPS for embedding search, 500 QPS for agent inference
Let’s explore your high-impact AI opportunities together in a complimentary session
