The Right Expert, Resource, or Event – Instantly: How Avahi Built an Al-Powered Discovery Platform for The Bloom

Client

The Bloom

Location

Pittstown, New Jersey

Industry

Social Impact / Community Development

Services & Tech

Amazon Bedrock Claude 3.5 Sonnet Claude 3 Haiku Amazon Titan Embeddings Amazon OpenSearch Serverless AWS Lambda Amazon S3 Amazon EventBridge Amazon CloudWatch AWS IAM Next.js

Client

The Bloom

Location

Pittstown, New Jersey

Industry

Social Impact / Community Development

Services & Tech

Amazon Bedrock Claude 3.5 Sonnet Claude 3 Haiku Amazon Titan Embeddings Amazon OpenSearch Serverless AWS Lambda Amazon S3 Amazon EventBridge Amazon CloudWatch AWS IAM Next.js

Project Overview

The Bloom is a social impact organization that connects community members with experts, educational resources, and events across areas like climate finance, social justice, and community development. As their platform and content library scaled, a rigid keyword search engine made it increasingly difficult for members to find what they actually needed, stalling the connections the platform was built to create. Avahi designed and deployed a fully serverless, Alpowered search and discovery platform on AWS, combining semantic search, RAG-powered natural language query interpretation, and an automated event indexing pipeline. The result is an intelligent, self-maintaining community platform that keeps itself current and accurate, with no ongoing operational overhead.

About The
 Customer

The Bloom is a Pittstown, New Jersey-based social impact organization dedicated to connecting community members with the people, knowledge, and opportunities relevant to their work. Operating across focus areas including climate finance, social justice, and community development, The Bloom runs a curated platform that brings together expert practitioners, educational content, and events for a mission-driven community. As a lean organization in the social impact space, The Bloom depends on its platform’s ability to make meaningful connections efficiently and at a sustainable cost.

The 
Problem

The Bloom’s core value proposition is connection: putting the right expert, resource, or event in front of the right community member at the right time. As theplatform’s library of profiles, documents, articles, reports, and events grew, that promise became harder to keep. A traditional keyword search engine – matching terms against text fields – is too rigid to handle the way people in a specialized, mission-driven community actually search. A query like “Who can I speak to about climate finance?” is not a keyword lookup. It is an expression of intent, context, and need that requires semantic understanding to answer well.

The gap between what community members were asking and what keyword search could surface was a discoverability problem with compounding consequences. Members who couldn’t find relevant experts or resources disengaged.

Events went unattended not because they lacked relevance, but because they were invisible to the people who would have valued them most. For a platform whose success depends entirely on facilitating high-quality connections, a search experience that couldn’t keep pace with a growing, diverse content library was a direct threat to the organization’s mission and member retention.

The events challenge added a further operational dimension. Keeping an event index accurate and current required manual effort, and for a lean social impact organization, that overhead was not scalable. Without automation, the event discovery experience would always lag behind reality, eroding trust in the platform’s usefulness over time.

Why AWS

AWS offered the combination of managed Al services, serverless compute, and vector search infrastructure that made it possible to build a sophisticated discovery platform without the cost and complexity of custom model deployment or self-managed infrastructure. Amazon Bedrock provided direct access to frontier language models – Claude 3.5 Sonnet, Claude 3 Haiku, and Titan Embeddings – through a single, governed API, eliminating the need to manage model hosting while preserving the ability to implement automatic fallback logic across models. This was particularly important for an organization like The Bloom, where service reliability and operational simplicity are both essential.

Amazon OpenSearch Serverless provided a fully managed vector search layer capable of running hybrid queries – combining traditional keyword scoring with KNN vector similarity – without requiring The Bloom to provision or maintain search cluster infrastructure. Paired with AWS Lambda, Amazon EventBridge, and S3 event notifications, the full pipeline runs on a serverless architecture that scales to demand and keeps operational costs minimal. For a social impact organization operating lean, an estimated monthly pipeline cost of $15-20 made AWS the only platform that could deliver this level of Al capability at a sustainable price point.

Why The Bloom Chose Avahi

The Bloom needed more than a vendor who could configure AWS services. They needed a partner who understood how to design an Al architecture that was simultaneously powerful, reliable, and operationally sustainable for a mission-driven organization with limited engineering resources. That combination of technical sophistication and practical judgment is what Avahi brought to the engagement.

Avahi’s approach to resilience engineering was a differentiating factor from the outset. Rather than building a pipeline that depended on any single external source or model behaving predictably, Avahi designed layered redundancy into every critical workflow, a web content extraction, event scraping, and Al model calls alike. This meant The Bloom would geta platform that degraded gracefully under real-world conditions rather than failing outright when an external site blocked a crawler or a model hit a rate limit.

Avahi also brought the experience to architect for client independence. Full technical documentation, a deployment guide, and a clean serverless architecture meant The Bloom’s team could maintain and extend the platform without ongoing external dependency, a critical consideration for an organization that needed a durable solution, not a managed service relationship.

Solution

Avahi built a fully serverless Al-powered discovery platform on AWS, operating across threе interconnected search and ingestion workflows, all unified in a single interface accessible through a Next.js frontend.

  • Expert and Resource Search When a community member submits a query, the Next.js APІ layer routes the request to Amazon Bedrock’s Titan Embeddings model, which generates a 1,536-dimensional vector representation of the query text. That vector is then passed to Amazon OpenSearch Serverless, which executes a hybrid search across the expert and resource indices, combining traditional keyword matching with KNN vector similarity using the HNSW algorithm and cosine similarity scoring. Results are filtered by relevance score thresholds and returned as interactive cards, surfacing semantically relevant experts and resources even when the query doesn’t match exact field values.
  • RAG-Powered Event Search Event discovery uses a distinct, two-stage approach. Claude 3 Haiku first interprets the natural language query, extracting location, theme, and keywords at temperature 0 for deterministic, structured output. The system then selects the optimal search strategy (location-only, theme-only, or combined), generates a Titan embedding for the theme component, and executes a KNN vector search with geographic filters against the events index in OpenSearch, returning up to 20 upcoming events ranked by semantic and geographic relevance.
  • Automated Event Pipeline The most operationally significant component of the platform is its fully automated event discovery pipeline. Amazon EventBridge triggers a Python Lambda function each morning on a scheduled basis. That function scrapes The Bloom’s Luma calendar using four sequential extraction strategies – browser-simulated HTTP requests, JSON-LD structured data parsing, Next.js embedded data extraction, and direct Luma API calls – with each fallback triggered automatically if the preceding method fails. Extracted events are saved as CSV to Amazon S3, which fires an S3 event notification that automatically triggers a second Lambda function to generate Titan embeddings for each event and bulk-index them into OpenSearch. From that point, events are immediately searchable, with no human intervention required after the initial setup.
  • Document Ingestion Workflow For content ingestion, Avahi built a multi-strategy web extraction pipeline that accepts a URL and attempts content retrieval using four sequential methods: direct HTTP fetch with user agent rotation, Puppeteer API for JavaScript-rendered pages, proxy service routing, and URL structure analysis as a final fallback. Once content is extracted, Claude 3.5 Sonnet analyzes it to generate a structured title, a standardized 100- word description, and category classifications, with automatic fallback to Claude 3 Sonnet and then Claude 3 Haiku if the primary model is unavailable. Titan Embeddings then generates a vector for the resource, which is appended to a shared CSV in S3 and made searchable through the same OpenSearch hybrid index. Amazon CloudWatch provides logging and performance monitoring across all Lambda functions, and AWS IAM enforces fine-grained access control across the full service architecture.

Key Deliverables

  • Serverless hybrid semantic search platform covering three content types: experts, resources, and events
  • RAG-powered event search with natural language query interpretation via Claude 3 Haiku
  • Automated event discovery and indexing pipeline (EventBridge + dual Lambda functions + S3 event notifications)
  • Multi-strategy web content ingestion workflow with four-layer fallback extraction and Claude 3.5 Sonnet analysis
  • Automatic Al model fallback cascade (Claude 3.5 Sonnet → Claude 3 Sonnet → Claude 3 Haiku)
  • Amazon OpenSearch Serverless indices for experts, resources, and events with hybrid KNN + keyword scoring
  • Streamlit UI for platform testing
  • RESTful API endpoints for all search and ingestion functions
  • Full technical documentation covering architecture, configuration, troubleshooting, and maintenance
  • AWS deployment guide covering Lambda, OpenSearch Serverless, S3, Bedrock, and EventBridge

Project
 Impact

Avahi delivered a fully deployed, self-maintaining Al discovery platform that fundamentally changes how The Bloom’s community members connect with experts, resources, and events. Natural language queries — the way people actually search – now return semantically relevant results across all three content types, replacing the rigid limitations of keyword matching with genuine intent-based discovery. The automated event pipeline eliminates the manual effort that previously stood between The Bloom’s Luma calendar and a searchable, up-to-date event index.

For a lean social impact organization, the architecture’s cost profile is as significant as its capabilities. The platform operates on fully managed, serverless AWS infrastructure – purposebuilt to scale without operational overhead and priced to remain sustainable long after the engagement closed.

Outcome Highlights

  • Natural language search enabled across three content types, experts, resources, and events, in a single unified interface
  • Estimated automated event pipeline cost: ~$15-20/month on serverless AWS infrastructure
  • Event index updated daily and automatically with zero manual intervention after initial setup
  • Four-layer fallback web extraction strategy ensures content ingestion from a wide range of external sources without manual intervention
  • Four-method fallback Luma event scraper maintains pipeline continuity even when primary extraction methods are blocked or unavailable
  • Automatic Al model fallback cascade across three Claude models maintains service availability during model outages or rate limiting events
  • KNN vector search with geographic filtering returns up to 20 semantically and geographically ranked upcoming events per query
  • Titan Embeddings generates 1,536-dimensional vectors for all experts, resources, events, and user queries powering platform-wide semantic search

Ready to Transform Your Business with AI?

Let’s explore your high-impact AI opportunities together in a complimentary session

Ready to Transform Your Business with AI?

Let’s explore your high-impact AI opportunities together in a complimentary session

Ready to Transform Your Business with AI?

Book Your Free Ignition AI Workshop

Let’s explore your high-impact AI opportunities together in a complimentary half-day session

View Our Case Studies

See how we’ve delivered measurable results for businesses like yours