Lucrum
Technology

How Lucrum works under the hood.

A multi-agent AI system built on AWS, designed for personal, safe and fast conversations at scale.

  • AWS
  • Amazon Bedrock
  • Amazon OpenSearch
  • LangGraph
  • NeMo Guardrails
  • Llama
  • Mistral
01

Architecture

AppiOS, Android, webAPI layerAWSSafety checkNeMo GuardrailsOrchestrationLangGraphLong-term memoryAmazon OpenSearch vector indexLanguage modelsBedrock + fine-tuned open modelsAnalytics and evaluationSafety checkbefore the replyREPLY

Every message passes through the safety layer before and after the model responds.

02

Specialized agents, one conversation

Onboarding agent

Runs the first conversation, captures goals and routine, and sets the starting plan.

Conversation agent

Handles everyday chat, using memory and recent check-ins for context.

Program agent

Builds and adjusts programs based on progress, mood and schedule.

Insights agent

Turns check-in data into weekly patterns explained in plain language.

03

The right model for each job

Production conversations run on foundation models through Amazon Bedrock. In parallel we fine-tune open models (Llama and Mistral, 7B to 13B parameters) on curated wellbeing dialogue for tone, empathy and safety, and route each task to the model with the best quality, speed and cost.

One taskIllustrative
Foundation modelAmazon Bedrock
Quality
Speed
Cost
Fine-tuned Llama7B to 13B
Quality
Speed
Cost
Fine-tuned Mistral7B to 13B
Quality
Speed
Cost
04

Personal, private memory

Key facts from conversations are summarized and stored as embeddings in an Amazon OpenSearch vector index. The companion retrieves only what is relevant to the current conversation. Users can view and delete their memory at any time.

Conversation

How did the presentation on Thursday go?

Better than expected. The walk before helped.

Summarized
Presentation on ThursdayWalks lift your mood
Amazon OpenSearch vector indexViewDelete
05

Safety pipeline

  1. 01
    Message in
    From the app
  2. 02
    Safety check
    NeMo Guardrails + crisis detection
  3. 03
    Model
    Orchestrated by LangGraph
  4. 04
    Safety check out
    Before the reply

During the beta, our team reviews every flagged conversation to improve how Lucrum responds.

06

Built on AWS

LayerWhat we useWhy
ModelsAmazon Bedrock; fine-tuned Llama and Mistral (7B to 13B parameters)Managed access to foundation models; open models tuned for tone, empathy and safety
OrchestrationLangGraphCoordinates the specialized agents in one conversation
MemoryAmazon OpenSearch ServiceVector memory and search
ComputeGPU instancesFine-tuning and low-latency inference
SafetyNeMo GuardrailsChecks every message before and after the model responds
DataHIPAA-eligible servicesEncryption by default; infrastructure that scales from beta to public launch
07

Our next 12 months

  • Fine-tune and evaluate our own wellbeing models on GPU instances
  • Cut response latency for a natural chat feel
  • Expand safety evaluation with red-teaming and human review
  • Launch voice check-ins