Production-ready cloud-native AI platform delivered end to end
Response latency, down from roughly ten seconds
Serverless stack idles at near-zero cost between sessions
United Sales
Palo Alto, California
AI-Powered Sales Productivity / Sales Technology
Amazon Bedrock (Claude), Amazon API Gateway, AWS Lambda, Amazon Cognito, Amazon OpenSearch Service, Amazon S3, Amazon DynamoDB, AWS Comprehend, AWS Textract, Amazon EventBridge, AWS VPC, AWS PrivateLink, AWS IAM, Amazon CloudWatch, AWS Secrets Manager, Chrome Extension (Manifest v3)
United Sales had built a browser-based AI assistant called Stella to help sales representatives answer technical product questions live during customer calls. The initial version ran entirely on a single machine, relied on a third-party AI provider through a hardcoded API credential, and had no persistent knowledge base or cloud backend. Avahi re-architected and delivered Stella as a production-ready, cloud-native platform on AWS within a five-week engagement. The result is a fully deployed AI sales copilot that ingests product documentation, processes live meeting captions, detects deal signals in real time, and surfaces coaching guidance to sales representatives during calls, without requiring a technical expert to be present.
United Sales is an AI-powered sales productivity company whose platform is designed to eliminate the dependency on sales engineers during customer conversations. The product serves sales teams that sell technically complex products, bridging the gap between sales expertise and deep technical knowledge by enabling Stella to answer customer questions accurately and in real time. United Sales operates as an early-stage technology company building toward enterprise customer pilots.
Sales representatives at technically complex companies routinely encounter questions they cannot answer without a subject-matter expert. When a sales engineer is unavailable, deals stall or collapse. United Sales had already validated this problem and built a working prototype, a Chrome extension called Stella that used a local AI model to answer questions during calls. But the prototype had critical limitations that blocked commercial scale.
Stella was bound to a single device, had no persistent product knowledge base, and required manual re-upload of reference documents for every new session. The system used a hardcoded provider API credential with no authentication layer, creating a security exposure. There was no mechanism to capture live meeting signals, no session history, and no way to track what was discussed across calls. The architecture could not support multiple sales representatives or serve as the foundation for investor demonstrations or early customer pilots.
United Sales’ founders recognized early that their target enterprise customers almost universally operate on AWS and rely on Claude as their preferred AI model. Migrating Stella onto Amazon Web Services with Amazon Bedrock as the inference layer meant the product could eventually run as an extension within a customer’s existing cloud environment, making token costs customer-borne rather than United Sales-borne. This commercial architecture decision shaped every technical choice: AWS Lambda for serverless, pay-per-use compute; Amazon Bedrock Knowledge Bases with OpenSearch for persistent, per-representative document retrieval; Amazon Cognito for authentication; and AWS Comprehend for real-time PII masking before any inference call. The entire stack idles at near-zero cost between active sessions, a critical constraint for an early-stage company managing a lean cloud budget.
United Sales needed a partner who could move quickly, work within tight cost constraints, and deliver a production-deployable result rather than a slide deck. Avahi’s standing as a Premier-tier AWS partner, combined with a structured five-week delivery model with weekly gate sign-offs, gave the founding team confidence that the engagement would produce a transferable, independently operable system. The team’s experience designing Bedrock-native architectures for AI-intensive workloads aligned directly with the technical scope.
Avahi rebuilt Stella from the ground up on AWS while preserving the Chrome extension as the user-facing interface. The delivered system consists of four integrated layers.
The knowledge layer uses an automated S3-to-Bedrock Knowledge Base synchronization pipeline. When a sales representative uploads a product document, an event-driven pipeline immediately indexes it into an OpenSearch k-nearest-neighbor semantic search index. Documents are scoped per sales representative so each rep retrieves only their own product materials, and the index persists across sessions, eliminating the manual re-upload requirement. AWS Textract handles extraction from PDF, Word, PowerPoint, and Excel formats.
The inference layer replaces the original third-party API dependency with Amazon Bedrock running Claude. The Chrome extension sends live caption text and the representative’s question to an Amazon API Gateway endpoint. A Lambda function retrieves the most relevant document passages from the knowledge base using semantic search, constructs a grounded prompt, and streams the response back to the extension via server-sent events. Response latency was reduced to three to five seconds through streaming, down from approximately ten seconds in the non-streaming prototype.
The signal detection layer processes live meeting captions through a taxonomy of deal signals developed in collaboration with United Sales’ sales and product advisors. The system identifies events such as budget objections, competitor mentions, quantified business pain, timeline pressure, and affirmation or contradiction patterns. Signals are ranked by priority and surfaced to the representative as contextual nudges, with the cited transcript line as evidence. The taxonomy went through four revision cycles informed by input from United Sales’ advisory network of sales professionals.
The security and data layer replaced the hardcoded API credential with Amazon Cognito token-based authentication across all endpoints. All traffic routes through a Virtual Private Cloud with AWS PrivateLink, keeping Bedrock and OpenSearch calls entirely within the AWS network. AWS Comprehend scans every caption batch and masks names, email addresses, phone numbers, and other personally identifiable information before any data reaches the inference layer. Session transcripts are stored in DynamoDB for the active period and archived to Amazon S3, giving the platform an audit trail that supports future coaching analytics and win-loss analysis.
The final week of the engagement covered user acceptance testing against live sales call scenarios, migration of the full stack into the United Sales AWS account, and knowledge transfer sessions covering the architecture, deployment pipeline, and day-to-day operations.
Stella codebase assessment: full inventory of source files, migration risks, and Gemini-to-Bedrock dependency mapping
AWS architecture: Lambda-based serverless stack with API Gateway, Cognito, VPC and PrivateLink, DynamoDB, S3, and Amazon Bedrock Knowledge Bases
API contract: authenticated endpoints for real-time query, document ingestion, caption processing, and signal detection
Automated S3-to-Bedrock Knowledge Base sync pipeline with AWS Textract for multi-format document support
RAG retrieval pipeline with OpenSearch k-NN semantic search and per-representative document scoping
Chrome extension migration from Gemini API to AWS API Gateway with Cognito authentication
Real-time signal detection engine based on the finalized version 1.4 deal signal taxonomy
PII masking via AWS Comprehend applied to all caption and query data before inference
Session history migrated from browser local storage to DynamoDB with S3 archival
Production deployment to the United Sales AWS account
Full technical documentation, architecture diagrams, and knowledge transfer sessions
The engagement delivered a fully deployed, independently operable sales AI platform. United Sales’ team can ingest product documentation, conduct live sales calls with real-time AI assistance, review detected deal signals, and manage the AWS environment without ongoing Avahi involvement. The platform is positioned for immediate pilot customer onboarding.
Response latency reduced from approximately 10 seconds to 3 to 5 seconds through streaming inference, meeting the sub-six-second latency requirement defined in the statement of work
Five formal phase sign-offs completed across the engagement, confirming all deliverables met agreed requirements before each successive phase began
Security posture upgraded from a hardcoded API credential to Cognito token authentication, VPC-internal routing, and PII masking applied to every caption and query payload
Knowledge base now supports persistent, multi-session document retrieval per sales representative, replacing a session-scoped system that required re-upload for every call
Deal signal taxonomy finalized after four revision cycles with input from sales practitioners and subject-matter advisors
Full stack deployed to production in the United Sales AWS account with source code, documentation, and runbooks transferred at engagement close
United Sales
Palo Alto, California
AI-Powered Sales Productivity / Sales Technology
Amazon Bedrock (Claude), Amazon API Gateway, AWS Lambda, Amazon Cognito, Amazon OpenSearch Service, Amazon S3, Amazon DynamoDB, AWS Comprehend, AWS Textract, Amazon EventBridge, AWS VPC, AWS PrivateLink, AWS IAM, Amazon CloudWatch, AWS Secrets Manager, Chrome Extension (Manifest v3)
United Sales had built a browser-based AI assistant called Stella to help sales representatives answer technical product questions live during customer calls. The initial version ran entirely on a single machine, relied on a third-party AI provider through a hardcoded API credential, and had no persistent knowledge base or cloud backend. Avahi re-architected and delivered Stella as a production-ready, cloud-native platform on AWS within a five-week engagement. The result is a fully deployed AI sales copilot that ingests product documentation, processes live meeting captions, detects deal signals in real time, and surfaces coaching guidance to sales representatives during calls, without requiring a technical expert to be present.
United Sales is an AI-powered sales productivity company whose platform is designed to eliminate the dependency on sales engineers during customer conversations. The product serves sales teams that sell technically complex products, bridging the gap between sales expertise and deep technical knowledge by enabling Stella to answer customer questions accurately and in real time. United Sales operates as an early-stage technology company building toward enterprise customer pilots.
Sales representatives at technically complex companies routinely encounter questions they cannot answer without a subject-matter expert. When a sales engineer is unavailable, deals stall or collapse. United Sales had already validated this problem and built a working prototype, a Chrome extension called Stella that used a local AI model to answer questions during calls. But the prototype had critical limitations that blocked commercial scale.
Stella was bound to a single device, had no persistent product knowledge base, and required manual re-upload of reference documents for every new session. The system used a hardcoded provider API credential with no authentication layer, creating a security exposure. There was no mechanism to capture live meeting signals, no session history, and no way to track what was discussed across calls. The architecture could not support multiple sales representatives or serve as the foundation for investor demonstrations or early customer pilots.
United Sales’ founders recognized early that their target enterprise customers almost universally operate on AWS and rely on Claude as their preferred AI model. Migrating Stella onto Amazon Web Services with Amazon Bedrock as the inference layer meant the product could eventually run as an extension within a customer’s existing cloud environment, making token costs customer-borne rather than United Sales-borne. This commercial architecture decision shaped every technical choice: AWS Lambda for serverless, pay-per-use compute; Amazon Bedrock Knowledge Bases with OpenSearch for persistent, per-representative document retrieval; Amazon Cognito for authentication; and AWS Comprehend for real-time PII masking before any inference call. The entire stack idles at near-zero cost between active sessions, a critical constraint for an early-stage company managing a lean cloud budget.
United Sales needed a partner who could move quickly, work within tight cost constraints, and deliver a production-deployable result rather than a slide deck. Avahi’s standing as a Premier-tier AWS partner, combined with a structured five-week delivery model with weekly gate sign-offs, gave the founding team confidence that the engagement would produce a transferable, independently operable system. The team’s experience designing Bedrock-native architectures for AI-intensive workloads aligned directly with the technical scope.
Avahi rebuilt Stella from the ground up on AWS while preserving the Chrome extension as the user-facing interface. The delivered system consists of four integrated layers.
The knowledge layer uses an automated S3-to-Bedrock Knowledge Base synchronization pipeline. When a sales representative uploads a product document, an event-driven pipeline immediately indexes it into an OpenSearch k-nearest-neighbor semantic search index. Documents are scoped per sales representative so each rep retrieves only their own product materials, and the index persists across sessions, eliminating the manual re-upload requirement. AWS Textract handles extraction from PDF, Word, PowerPoint, and Excel formats.
The inference layer replaces the original third-party API dependency with Amazon Bedrock running Claude. The Chrome extension sends live caption text and the representative’s question to an Amazon API Gateway endpoint. A Lambda function retrieves the most relevant document passages from the knowledge base using semantic search, constructs a grounded prompt, and streams the response back to the extension via server-sent events. Response latency was reduced to three to five seconds through streaming, down from approximately ten seconds in the non-streaming prototype.
The signal detection layer processes live meeting captions through a taxonomy of deal signals developed in collaboration with United Sales’ sales and product advisors. The system identifies events such as budget objections, competitor mentions, quantified business pain, timeline pressure, and affirmation or contradiction patterns. Signals are ranked by priority and surfaced to the representative as contextual nudges, with the cited transcript line as evidence. The taxonomy went through four revision cycles informed by input from United Sales’ advisory network of sales professionals.
The security and data layer replaced the hardcoded API credential with Amazon Cognito token-based authentication across all endpoints. All traffic routes through a Virtual Private Cloud with AWS PrivateLink, keeping Bedrock and OpenSearch calls entirely within the AWS network. AWS Comprehend scans every caption batch and masks names, email addresses, phone numbers, and other personally identifiable information before any data reaches the inference layer. Session transcripts are stored in DynamoDB for the active period and archived to Amazon S3, giving the platform an audit trail that supports future coaching analytics and win-loss analysis.
The final week of the engagement covered user acceptance testing against live sales call scenarios, migration of the full stack into the United Sales AWS account, and knowledge transfer sessions covering the architecture, deployment pipeline, and day-to-day operations.
Stella codebase assessment: full inventory of source files, migration risks, and Gemini-to-Bedrock dependency mapping
AWS architecture: Lambda-based serverless stack with API Gateway, Cognito, VPC and PrivateLink, DynamoDB, S3, and Amazon Bedrock Knowledge Bases
API contract: authenticated endpoints for real-time query, document ingestion, caption processing, and signal detection
Automated S3-to-Bedrock Knowledge Base sync pipeline with AWS Textract for multi-format document support
RAG retrieval pipeline with OpenSearch k-NN semantic search and per-representative document scoping
Chrome extension migration from Gemini API to AWS API Gateway with Cognito authentication
Real-time signal detection engine based on the finalized version 1.4 deal signal taxonomy
PII masking via AWS Comprehend applied to all caption and query data before inference
Session history migrated from browser local storage to DynamoDB with S3 archival
Production deployment to the United Sales AWS account
Full technical documentation, architecture diagrams, and knowledge transfer sessions
The engagement delivered a fully deployed, independently operable sales AI platform. United Sales’ team can ingest product documentation, conduct live sales calls with real-time AI assistance, review detected deal signals, and manage the AWS environment without ongoing Avahi involvement. The platform is positioned for immediate pilot customer onboarding.
Response latency reduced from approximately 10 seconds to 3 to 5 seconds through streaming inference, meeting the sub-six-second latency requirement defined in the statement of work
Five formal phase sign-offs completed across the engagement, confirming all deliverables met agreed requirements before each successive phase began
Security posture upgraded from a hardcoded API credential to Cognito token authentication, VPC-internal routing, and PII masking applied to every caption and query payload
Knowledge base now supports persistent, multi-session document retrieval per sales representative, replacing a session-scoped system that required re-upload for every call
Deal signal taxonomy finalized after four revision cycles with input from sales practitioners and subject-matter advisors
Full stack deployed to production in the United Sales AWS account with source code, documentation, and runbooks transferred at engagement close
Let’s explore your high-impact AI opportunities together in a complimentary session
