PUMA Vault

🤖 Tools — AI Assistants & LLM Models

Se lee en 4 min

🤖 Tools — AI Assistants & LLM Models

Overview

PUMA uses AI models in two distinct roles:

  • Role A: Research and development assistance (Marco Veritas protocol applies)
  • Role B: Object of study — models evaluated in the benchmark experiment

Role A — Research & Development Assistance

Claude (Anthropic)

  • URL: https://claude.ai
  • Model: claude-sonnet-4-6 (primary), claude-opus-4-6 (deep analysis tasks)
  • Role: A — Research assistant and writing support
  • Phase: F0 – continuous
  • PUMA use:
    • Literature synthesis and state-of-the-art analysis
    • Argumentative coherence review of TFG chapters
    • Architectural decision analysis (BMAD Agent Architect role)
    • Prompt engineering for experiment strategies
    • Code scaffolding review (always verified by author)
  • Strengths: 200K token context; strong reasoning over long documents; best for complex synthesis tasks
  • Declaration: All uses declared in Section 1.8 (AI use declaration, Marco Veritas protocol)
  • Related: PT-Claude-RCOIF-Research

ChatGPT (OpenAI)

  • URL: https://chatgpt.com
  • Model: GPT-4o (text+code tasks)
  • Role: A — Conversational assistant and idea contrast
  • Phase: F0 – F2
  • PUMA use:
    • Contrasting ideas with a different system (avoiding single-model bias)
    • Exploring alternative architectural approaches
    • Draft generation for manual rewriting
  • Note: All output rewritten in author’s voice before inclusion (Marco Veritas)

Google Gemini (AI Studio)

  • URL: https://aistudio.google.com
  • Model: Gemini 1.5 Flash (free tier: 15 req/min)
  • Role: A (exploration) + B (experiment evaluation — see below)
  • Phase: F0–F1 (Role A), F4 (Role B)
  • Role A use: Bibliographic exploration; synthesis of PM ecosystem; complementary perspective to Claude/GPT-4o
  • Note: Free API tier used; rate limits noted in experiment design

DeepSeek-R1

  • URL: https://www.deepseek.com
  • Role: A (technical reasoning) + B (experiment candidate)
  • Phase: F1 – F4
  • PUMA use (Role A): Complex technical architecture problems; chain-of-thought reasoning tasks
  • PUMA use (Role B): Evaluated in experiment comparative; available via API and via Ollama locally
  • Strengths: Explicit reasoning traces; open-weights version available for local execution

Microsoft Copilot

  • URL: https://copilot.microsoft.com
  • Role: A — Microsoft environment assistant
  • Phase: F3 – F5
  • PUMA use: Writing assistance within Word/Teams environment; Spanish grammar checking for TFG sections

NotebookLM (Google)

  • URL: https://notebooklm.google.com
  • Role: A — RAG conversational interface over own documents
  • Phase: F0 – F1
  • PUMA use:
    • Building a private knowledge base from downloaded papers
    • Semantic queries without rereading (Keshav Pass-1 acceleration)
    • Audio synthesis of literature summaries for assimilation
  • Justification: Private document Q&A without external API; validated that responses match source content (Marco Veritas)
  • Related: Keshav-Reading-Log

Jenni AI

  • URL: https://jenni.ai
  • Role: A — Academic writing assistant
  • Phase: F3 – F5
  • PUMA use: Structuring formal academic sections; maintaining consistent academic style; citation suggestions (verified manually)

Role B — Experiment Models (Objects of Study)

Llama 3.2 8B (Meta) via Ollama

  • URL: https://ollama.ai/library/llama3.2
  • Role: B — Primary experiment model
  • Phase: F2 – F4
  • Quantisation: Q4_K_M (~5GB RAM)
  • Parameters: 8 billion
  • License: Meta Llama 3 Community License (open-weights)
  • PUMA use: PRIMARY model for Stage 1 (triage) and Stage 2 (estimation) experiments
  • Reproducibility settings: seed=42, temperature=0, num_predict=50 (triage) / 20 (estimation)
  • Install: ollama pull llama3.2:8b-instruct-q4_K_M

Mistral 7B via Ollama

  • URL: https://ollama.ai/library/mistral
  • Role: B — Comparison model
  • Phase: F3 – F4
  • Quantisation: Q4_K_M (~4.5GB RAM)
  • Parameters: 7 billion
  • License: Apache 2.0 (fully open)
  • PUMA use: SECONDARY model for comparison with Llama 3.2 8B; faster inference on CPU
  • Install: ollama pull mistral:7b-instruct-q4_K_M

Phi-3.5 Mini via Ollama

  • URL: https://ollama.ai/library/phi3.5
  • Role: B — Fallback model (latency mitigation)
  • Phase: F2 – F4
  • Quantisation: Q4_K_M (~2GB RAM)
  • Parameters: 3.8 billion
  • License: MIT
  • PUMA use: FALLBACK if Llama 3.2 8B latency > 60s on target hardware; maintains <30s/inference
  • Risk mitigation: Addresses latency risk (Gestión de riesgos, Section 1.5)
  • Install: ollama pull phi3.5:3.8b-mini-instruct-q4_K_M

Groq API (fast inference)

  • URL: https://groq.com
  • Role: B — High-speed inference option for comparative runs
  • Phase: F4
  • PUMA use: Running Llama 3.1 at hardware-accelerated speed (LPU inference); useful for comparative experiment requiring many rapid calls
  • Note: Free tier available; API-based (not local); used only for comparative runs, not primary experiment (reproducibility requires local Ollama)

Ollama — Local LLM Server

  • URL: https://ollama.ai
  • Function: Open-source tool for running LLMs locally via REST API
  • Phase: F2 – F4
  • PUMA use: CENTRAL infrastructure component; serves Llama 3.2 8B, Mistral 7B, Phi-3.5 Mini, nomic-embed-text for all experiments
  • API compatibility: OpenAI-compatible REST API at http://localhost:11434
  • Reproducibility: Model content-addressed registry ensures pinned versions
  • Key commands: 
    ollama pull llama3.2:8b-instruct-q4_K_M
ollama pull mistral:7b-instruct-q4_K_M
ollama pull nomic-embed-text
ollama serve  # starts API server
  • Related: LN-Tools-Ollama-ClaudeCode-OpenCode-BrowserOS

LM Studio

  • URL: https://lmstudio.ai
  • Function: GUI application for running local LLMs; OpenAI-compatible REST API at localhost:1234
  • Phase: F1 – F3
  • PUMA use: Model exploration without programming; rapid prompt prototyping; model quality comparison before Ollama integration
  • Fallback: Primary fallback if Ollama installation fails on target hardware
  • Note: Same API endpoint format; code runs unmodified by changing base URL

MOCs

Vista Gráfica

Retroenlaces