Reunião OSX 29/09

Meeting purpose / context

• A project review meeting (“Reunião OSX 29/09”) focused on explaining two AI-enabled products and how they will be packaged for team use via a curated handoff.
• The goal is to help the team use the outputs without exposing unnecessary private code or materials.

Project management / sprint setup

• The speaker describes a plan/schedule project that tracks created “scruns” (tasks).
• The plan includes:
  • Status
  • A weekly review mechanism
• Key configuration emphasis:
  • Priority
  • Start date
  • A maximum execution time
• A guided sprint is planned, using an assistant/reference code as the starting point.
• Action items include:
  • Ensuring tasks are assigned
  • Ensuring dates are included (from today through the next step)

Core technological concept: RAG + SLM architecture (instead of direct cloud LLM calls)

• The prototypes aim for an architecture closer to “SLM + RAG” (not fully complete yet).
• The system builds a local/own database (“corpus”), then uses an assistant to query and operate over that data.
• Stated benefit:
  • RAG + SLM reduces hallucination risk compared to relying on cloud LLMs that may “make things up.”
• Reliance on quality:
  • Reliability depends on document quality and curation
  • The pipeline-generated indicators/standards are positioned as trustworthy sources.

Two products / prototypes being explained

1) StepGraph (mobility + personas from telecom usage records)

• Uses a pipeline-based approach to transform telecom records into analytics and persona clusters.
• High-level pipeline:
  • Input: usage records / CDR-like traces (“usage records” discussed as large-scale per day)
  • Compute mobility routes/trajectories per subscriber and convert them into tensors
  • Apply anonymization to prevent individual-level mobility exposure
• Data representations:
  • Sparse tensors (avoid filling missing cells with zeros)
  • 3D tensors: subscribers × indicators/features × time slices/days
• Persona building / ML:
  • Persona discovery via clustering and ML models
    • Mentions clustering similar to “K+ / Caminhos+”
    • Mentions XGBoost for classification
  • Builds:
    • “fitness features”
    • “pivoting” behavior matrices to incorporate longitudinal history (turning time into aggregated indicators)
• Feature & indicator catalogs:
  • Outputs include an indicator catalog and feature catalog (spreadsheet-like artifacts) defining:
    • tensor dimensions
    • used features
    • persona definitions and classification criteria
  • Mentions a glossary/help guide included in the RAG corpus (definitions of terms and meanings)
• Explainability approaches:
  • Uses:
    • graph theory
    • probabilistic transition networks
      • BAISA/state-change probability networks
      • Markov chains to determine “getting worse/better”
  • Builds graph indicators including:
    • centrality
    • “origin/extension multilayer” concepts
• Additional computed indicators:
  • IPED: an integrated/weighted indicator of “experience in displacement,” aggregating multiple signals (e.g., quality/latency/data-use changes)
  • “Vamp”-style functions to estimate behaviors (e.g., addiction patterns / early-hour behavior)
  • Coverage/population estimation

2) NetGraft (ISP/network management insights from graph concepts)

• Presented as a separate prototype for managing ISP networks.
• The focus is a technical reference defining logical layers.
• Key graph concepts mentioned:
  • Centrality: removing a node breaks topology / identifies critical nodes
  • SPOF/spoff: described as a “last” knot/node with no alternative; associated with impacted rates
  • Blast radius: likely affected spread area from nodes/events
• The speaker emphasizes:
  • The intent is to explain layers and required concepts, not internal code.

System architecture / implementation approach

• Both products are described as pipeline-driven:
  • Pipelines behave like chained functions/components
  • A React/HTML application serves as a UI that calls pipelines as needed
• Notes on code generation:
  • Code was largely generated using AI/prompt engineering
    • Includes “Cloud Code” and AI-generated prompt/jargon/scaffolding
  • Minimal manual hand-coding of Python/React/JS
• Handoff packaging:
  • The team should receive an excerpted handoff:
    • enough code + documentation to develop
    • not the full private repo or sensitive internal assets
• Isolation (“cut”) plan:
  • Keep the assistant/SLM environment isolated from the pipeline environment
  • Preferably fork/project handoff so the assistant can query curated outputs safely

Review questions & how they will guide the handoff

• The speaker asks participants:
  • If technical questions exist regarding the specification document
  • Whether they need help understanding concepts in the documents versus seeing the project structure directly
• Clarification request:
  • Participants need access to the relevant code excerpts (Python/HTML/etc.) required to develop the project
  • They should not require access to the full internal system

Deliverables / packaging plan

• The speaker plans to:
  • Produce a curated document and send it to the group in a few days
  • Provide a repository-like handoff including:
    • necessary pipeline components
    • the assistant setup (SLM/RAG)
    • curated corpus content (standards/context descriptors/glossary)
• Deployment considerations:
  • Mentions discussion around client-side/server deployment
  • Intends to coordinate further with Paulo on how to structure the handoff cleanly

Main speakers / sources (as implied by subtitles)

• Guto (Gustavo?) – participant asking/confirming points about structure and integration
• Paulo – referenced expert supporting deeper technical integration
• Claudia – referenced as someone who can help assemble/prepare parts of the corpus/data
• CTO / “Cto” – participant asking about “structures” and expecting code excerpts
• Main presenter/speaker (unnamed in subtitles) – lead technical explainer describing:
  • StepGraph / NetGraft
  • the RAG + SLM approach
  • the pipeline/tensor/graph/ML methods