Back to Projects

DreamoreAI

Production

10 agents. 90.2% improvement. Zero fine-tuning.

B2B human capital platform with 10-agent LangGraph orchestration, pgvector matching, and graph neural network scoring. Voice analysis, skill extraction, portfolio generation, company diagnostics, and talent matching — each handled by a specialized agent.

Screenshot coming soon

The Problem

Traditional hiring and talent management relies on keyword matching and gut instinct. Companies waste millions on bad fits, and great candidates get overlooked because their resume doesn't have the right buzzwords. The disconnect between what companies need and what people offer is a pattern-matching problem that humans solve poorly at scale.

The Solution

Multi-Agent Orchestration

10-agent LangGraph system coordinating voice analysis, skill extraction, portfolio generation, company diagnostics, and talent matching with fault-tolerant communication patterns.

Vector Similarity Matching

pgvector-powered candidate-to-company matching combining skills (40%), culture fit (30%), personality (20%), and growth potential (10%) through graph neural network scoring.

Responsible AI Framework

Bias mitigation built into the architecture: protected characteristics removed from embeddings, fairness constraints enforced, human oversight layers at every decision point.

$50K-100K Assessment Tool

AI-powered company diagnostic that evaluates organizational health, identifies talent gaps, and generates actionable transformation roadmaps.

Tech Stack

Next.js(Framework)
TypeScript(Language)
Python(Language)
PostgreSQL(Database)
Supabase(Backend)
pgvector(Vector DB)
OpenAI APIs(AI)
Claude API(AI)
LangGraph(Orchestration)
AWS Lambda(Infrastructure)
API Gateway(Infrastructure)
S3(Storage)
Docker(DevOps)
Tailwind CSS(Styling)

Key Decisions

pgvector over Pinecone

ACID compliance + vectors in a single database at $0 marginal cost. When you're a startup, every dollar matters—and having transactions and vectors in the same DB eliminates an entire class of consistency bugs.

Claude Subagents over Fine-Tuned Models

90.2% improvement in assessment quality. Fine-tuning locks you into training data; multi-agent orchestration lets you compose capabilities and iterate independently.

Graph Neural Networks for Matching

Traditional cosine similarity misses relationships. GNN scoring captures how skills relate to roles, how culture dimensions interact, and how growth trajectories align with company needs.