/
Thumbprint Furniture
Virginia, USA
B2B SaaS, commercial furniture technology
Amazon Bedrock (Agents, Knowledge Bases), Anthropic Claude Haiku 4.5 via Amazon Bedrock, Amazon Titan Text Embeddings v2, Amazon Aurora Serverless v2 (PostgreSQL with pgvector), AWS Lambda, Amazon API Gateway, AWS Fargate, Amazon S3, AWS CloudFormation, Amazon CloudWatch
Thumbprint Furniture is a B2B SaaS platform that streamlines commercial furniture specification for dealers, designers, and distributors. Its core product automates space planning and specification, but end customers had no direct pathway to discover and purchase furniture from external suppliers through the assistant. Avahi delivered a standalone, agentic furniture shopping assistant on AWS that accepts natural language queries and optional inspiration images, retrieves matching products from a curated multi-supplier catalog, and returns a ranked shortlist with direct purchase links. The solution runs on a fully serverless AWS architecture built around Amazon Bedrock, Anthropic Claude Haiku 4.5, and a pgvector-backed knowledge base, giving Thumbprint a working consumer-grade experience it can validate with real users and evolve into production.
Thumbprint Furniture is a venture-backed startup founded in 2023, building a cloud-based platform that automates the design and specification of commercial furniture. It serves the commercial contract furniture ecosystem: dealers, designers, distributors, and manufacturers who rely on fragmented tools and lengthy manual workflows to specify and sell furniture packages. The company’s platform combines 3D visualization, real-time pricing, and automated bill-of-materials generation to compress a process that traditionally takes months into minutes.
While the core Thumbprint platform excels at space programming, specification, and documentation for professional users, it offered no direct pathway for an end customer to discover and purchase furniture from an approved set of external suppliers through natural conversation. Buyers describing what they wanted, whether a mid-century modern sofa under a specific budget or dining chairs sized for a particular alcove, had no way to translate that intent into a curated shortlist without leaving the platform and returning to fragmented supplier websites.
Without a solution, Thumbprint faced two related risks. It could not validate a consumer-grade buying experience on top of its specification platform, and it could not demonstrate a monetization pathway through supplier referral flows. Building that experience directly into the existing production platform would have carried significant integration cost and delayed customer validation. Thumbprint needed a way to prove the concept, gather real user signals, and de-risk a future integration decision.
Thumbprint chose AWS as the foundation for the assistant because it gave the team a single, managed environment for the full stack the workload required: multimodal foundation models via Amazon Bedrock, a managed knowledge base with native vector search, serverless compute, and hosted embeddings. Consolidating on AWS removed the operational overhead of stitching together separate model providers, vector databases, and orchestration layers, which was critical given the compressed timeline for the initial engagement.
The AWS platform also gave Thumbprint room to scale. The same architectural primitives that powered the initial build (Amazon Bedrock Agents, Amazon Bedrock Knowledge Bases, Amazon Aurora Serverless v2, AWS Lambda, and Amazon API Gateway) support production workloads without a re-platforming step, making it a durable foundation for the roadmap beyond the initial engagement.
As a premier tier AWS partner with a focused GenAI practice, Avahi brought hands-on expertise in the Amazon Bedrock and serverless services this build required. Thumbprint needed a partner that could design and ship a working, well-architected system in weeks rather than months, and one that could translate an ambitious agentic-shopping concept into concrete architectural decisions without re-litigating the fundamentals.
Beyond the technical fit, Avahi offered structured project management, a consistent delivery cadence, and access to AWS funding that accelerated the engagement’s start. Thumbprint valued the ability to move from signed statement of work to a live, testable application inside the AWS environment with a partner that could pragmatically scope trade-offs as delivery progressed.
Avahi designed and delivered a fully serverless, GenAI-powered furniture shopping assistant on AWS. The application presents a split-screen web interface: a chat panel on the left, where users describe what they are looking for in natural language, and a results panel on the right that surfaces ranked product recommendations alongside a persistent “My Selection” list. Users can attach an inspiration image, adjust preferences mid-conversation, and drill into any recommendation with a direct link to the supplier’s product page.
At the core of the solution is an Amazon Bedrock Agent running on Anthropic Claude Haiku 4.5, which handles multimodal input, multi-turn context, clarification logic, and orchestration. When the user submits a query, a lightweight Chat Proxy Lambda validates the request, injects the current cart contents into Bedrock session state, forwards any inspiration image as inline byte content, and hands off to the agent. The agent maintains a running understanding of user preferences (style, budget, room, size constraints) across the conversation and surfaces a brief session summary when those preferences change.
Retrieval and ranking are handled by a single Lambda registered as a Bedrock action group. It runs a deterministic pipeline on every invocation. A Query Enricher merges the user’s text with style tags extracted from any inspiration image and relevant session memory. An Intent Parser decomposes the enriched query into a category filter, a semantic search input, a budget constraint, and dimension constraints. A hybrid search then runs against Amazon Bedrock Knowledge Bases, combining semantic vector similarity on Amazon Titan Text Embeddings v2 with hard metadata filtering on category. We moved towards a “local” budget (per item total)per customers preference and a weighted ranker produces the final ordered list.
The knowledge base is backed by Amazon Aurora Serverless v2 with pgvector, chosen deliberately over alternative vector stores. Aurora Serverless v2 can auto-pause to zero capacity units when idle, keeping costs near zero between sessions, while still delivering hybrid semantic and metadata search at query time. Product data is populated by an ingest pipeline: AWS Fargate tasks scrape approved supplier catalogs, extract structured product data, download and cache images to Amazon S3, and normalize categories and prices. A sidecar generator produces the embedding documents and metadata that Amazon Bedrock Knowledge Bases ingests. The ingest layer is deliberately isolated behind an adapter contract, so supplier API integrations can replace the scraper in a production deployment without touching the rest of the pipeline.
The frontend is a vanilla JavaScript single-page application hosted on Amazon S3, calling Amazon API Gateway which fronts the Chat Proxy Lambda. Runtime configuration is externalized to a config file colocated with the static assets, so the same build deploys to any environment. Infrastructure is defined in AWS CloudFormation across two stacks, deployed by GitHub Actions using OpenID Connect to assume an IAM deploy role, eliminating any long-lived AWS access keys in the CI/CD pipeline.
The retrieval pipeline uses a generalized tool contract, so new capabilities such as a lead-time filter, a style compatibility scorer, or a previously-viewed-items re-ranker can be added without modifying the pipeline orchestrator or any existing tool. That extensibility is a deliberate design choice to make the transition from an initial build to a production system additive rather than disruptive.
The engagement delivered a working, deployed shopping assistant on AWS, giving Thumbprint a hands-on demonstration of an agentic buying experience it could put in front of real users. The solution’s architectural decisions position Thumbprint for a low-friction path to production: the adapter layer in the ingest pipeline means moving from scraping to supplier APIs requires no changes downstream, auto-pause on Aurora Serverless v2 keeps idle costs near zero between sessions, the generalized tool contract in the retrieval pipeline means new ranking signals can be added without disrupting existing behavior, and CI/CD via GitHub Actions OpenID Connect eliminates long-lived credentials from the deploy path.
The client signed off the engagement with positive sentiment and is actively scoping a follow-on production phase.
Thumbprint Furniture
Virginia, USA
B2B SaaS, commercial furniture technology
Amazon Bedrock (Agents, Knowledge Bases), Anthropic Claude Haiku 4.5 via Amazon Bedrock, Amazon Titan Text Embeddings v2, Amazon Aurora Serverless v2 (PostgreSQL with pgvector), AWS Lambda, Amazon API Gateway, AWS Fargate, Amazon S3, AWS CloudFormation, Amazon CloudWatch
Thumbprint Furniture is a B2B SaaS platform that streamlines commercial furniture specification for dealers, designers, and distributors. Its core product automates space planning and specification, but end customers had no direct pathway to discover and purchase furniture from external suppliers through the assistant. Avahi delivered a standalone, agentic furniture shopping assistant on AWS that accepts natural language queries and optional inspiration images, retrieves matching products from a curated multi-supplier catalog, and returns a ranked shortlist with direct purchase links. The solution runs on a fully serverless AWS architecture built around Amazon Bedrock, Anthropic Claude Haiku 4.5, and a pgvector-backed knowledge base, giving Thumbprint a working consumer-grade experience it can validate with real users and evolve into production.
Thumbprint Furniture is a venture-backed startup founded in 2023, building a cloud-based platform that automates the design and specification of commercial furniture. It serves the commercial contract furniture ecosystem: dealers, designers, distributors, and manufacturers who rely on fragmented tools and lengthy manual workflows to specify and sell furniture packages. The company’s platform combines 3D visualization, real-time pricing, and automated bill-of-materials generation to compress a process that traditionally takes months into minutes.
While the core Thumbprint platform excels at space programming, specification, and documentation for professional users, it offered no direct pathway for an end customer to discover and purchase furniture from an approved set of external suppliers through natural conversation. Buyers describing what they wanted, whether a mid-century modern sofa under a specific budget or dining chairs sized for a particular alcove, had no way to translate that intent into a curated shortlist without leaving the platform and returning to fragmented supplier websites.
Without a solution, Thumbprint faced two related risks. It could not validate a consumer-grade buying experience on top of its specification platform, and it could not demonstrate a monetization pathway through supplier referral flows. Building that experience directly into the existing production platform would have carried significant integration cost and delayed customer validation. Thumbprint needed a way to prove the concept, gather real user signals, and de-risk a future integration decision.
Thumbprint chose AWS as the foundation for the assistant because it gave the team a single, managed environment for the full stack the workload required: multimodal foundation models via Amazon Bedrock, a managed knowledge base with native vector search, serverless compute, and hosted embeddings. Consolidating on AWS removed the operational overhead of stitching together separate model providers, vector databases, and orchestration layers, which was critical given the compressed timeline for the initial engagement.
The AWS platform also gave Thumbprint room to scale. The same architectural primitives that powered the initial build (Amazon Bedrock Agents, Amazon Bedrock Knowledge Bases, Amazon Aurora Serverless v2, AWS Lambda, and Amazon API Gateway) support production workloads without a re-platforming step, making it a durable foundation for the roadmap beyond the initial engagement.
As a premier tier AWS partner with a focused GenAI practice, Avahi brought hands-on expertise in the Amazon Bedrock and serverless services this build required. Thumbprint needed a partner that could design and ship a working, well-architected system in weeks rather than months, and one that could translate an ambitious agentic-shopping concept into concrete architectural decisions without re-litigating the fundamentals.
Beyond the technical fit, Avahi offered structured project management, a consistent delivery cadence, and access to AWS funding that accelerated the engagement’s start. Thumbprint valued the ability to move from signed statement of work to a live, testable application inside the AWS environment with a partner that could pragmatically scope trade-offs as delivery progressed.
Avahi designed and delivered a fully serverless, GenAI-powered furniture shopping assistant on AWS. The application presents a split-screen web interface: a chat panel on the left, where users describe what they are looking for in natural language, and a results panel on the right that surfaces ranked product recommendations alongside a persistent “My Selection” list. Users can attach an inspiration image, adjust preferences mid-conversation, and drill into any recommendation with a direct link to the supplier’s product page.
At the core of the solution is an Amazon Bedrock Agent running on Anthropic Claude Haiku 4.5, which handles multimodal input, multi-turn context, clarification logic, and orchestration. When the user submits a query, a lightweight Chat Proxy Lambda validates the request, injects the current cart contents into Bedrock session state, forwards any inspiration image as inline byte content, and hands off to the agent. The agent maintains a running understanding of user preferences (style, budget, room, size constraints) across the conversation and surfaces a brief session summary when those preferences change.
Retrieval and ranking are handled by a single Lambda registered as a Bedrock action group. It runs a deterministic pipeline on every invocation. A Query Enricher merges the user’s text with style tags extracted from any inspiration image and relevant session memory. An Intent Parser decomposes the enriched query into a category filter, a semantic search input, a budget constraint, and dimension constraints. A hybrid search then runs against Amazon Bedrock Knowledge Bases, combining semantic vector similarity on Amazon Titan Text Embeddings v2 with hard metadata filtering on category. We moved towards a “local” budget (per item total)per customers preference and a weighted ranker produces the final ordered list.
The knowledge base is backed by Amazon Aurora Serverless v2 with pgvector, chosen deliberately over alternative vector stores. Aurora Serverless v2 can auto-pause to zero capacity units when idle, keeping costs near zero between sessions, while still delivering hybrid semantic and metadata search at query time. Product data is populated by an ingest pipeline: AWS Fargate tasks scrape approved supplier catalogs, extract structured product data, download and cache images to Amazon S3, and normalize categories and prices. A sidecar generator produces the embedding documents and metadata that Amazon Bedrock Knowledge Bases ingests. The ingest layer is deliberately isolated behind an adapter contract, so supplier API integrations can replace the scraper in a production deployment without touching the rest of the pipeline.
The frontend is a vanilla JavaScript single-page application hosted on Amazon S3, calling Amazon API Gateway which fronts the Chat Proxy Lambda. Runtime configuration is externalized to a config file colocated with the static assets, so the same build deploys to any environment. Infrastructure is defined in AWS CloudFormation across two stacks, deployed by GitHub Actions using OpenID Connect to assume an IAM deploy role, eliminating any long-lived AWS access keys in the CI/CD pipeline.
The retrieval pipeline uses a generalized tool contract, so new capabilities such as a lead-time filter, a style compatibility scorer, or a previously-viewed-items re-ranker can be added without modifying the pipeline orchestrator or any existing tool. That extensibility is a deliberate design choice to make the transition from an initial build to a production system additive rather than disruptive.
The engagement delivered a working, deployed shopping assistant on AWS, giving Thumbprint a hands-on demonstration of an agentic buying experience it could put in front of real users. The solution’s architectural decisions position Thumbprint for a low-friction path to production: the adapter layer in the ingest pipeline means moving from scraping to supplier APIs requires no changes downstream, auto-pause on Aurora Serverless v2 keeps idle costs near zero between sessions, the generalized tool contract in the retrieval pipeline means new ranking signals can be added without disrupting existing behavior, and CI/CD via GitHub Actions OpenID Connect eliminates long-lived credentials from the deploy path.
The client signed off the engagement with positive sentiment and is actively scoping a follow-on production phase.
Let’s explore your high-impact AI opportunities together in a complimentary session
