LifeStack / MENTOR_PITCH.md
Soham Banerjee
deploy: pure lifestack with partitioned wisdom pool
77da5ce

Mentor Meeting Playbook — LifeStack Engine

The Core Framing

Research Question: "Can a small model (1.5B) learn to navigate multi-domain, causally-coupled crises better than a base LLM, using GRPO with a 7-day horizon reward?"


Slide Deck Structure (8 Slides Max)

Slide 1 — The Gap (30 sec)

  • Current AI: Single-turn advice, no state, no consequence modeling.
  • LifeStack: Life as a Markov Decision Process — 23 metrics, 6 domains, 40 causal edges.
  • Hook: "We built the environment that lets you train models on the 'ripple effects' of human decisions."

Slide 2 — The Environment (1 min)

  • Standards-Based: LifeStackEnv extends openenv.Environment.
  • Causal Foundation: 40 edges from Starcke & Brand (2012) — research-grounded, not arbitrary.
  • Deterministic World: DependencyGraph.propagate() uses matrix math, not LLM hallucination.
  • State Vector: 26-dim observation space across 23 tracked metrics.

Slide 3 — The Cascade (The Visual Hook)

  • Visual: Screenshot/GIF of the 4-frame cascade animation (STABLE → DISRUPTION → 1ST CASCADE → 2ND CASCADE).
  • Narrative: "A $350 flight rebooking cascades into stress (day 1) → sleep loss (day 2) → relationship strain (day 4). Our graph engine computes this propagation."

Slide 4 — Training Setup (45 sec)

  • Model: Qwen2.5-1.5B-Instruct, fine-tuned with GRPO via HuggingFace TRL.
  • Reward: 7-signal orchestrator (Milestone, Outcome, Preservation, Replan, Efficiency, Reasoning Coherence).
  • Innovation: $\gamma=0.9$ discounted 7-day rollout. Decisions are penalized today if they cause system collapse on day 4.

Slide 5 — The Research Result (Comparison)

Feature Untrained LLM (Base) GRPO-Trained LifeStack
Logic Treats each action independently Reasons across all 6 domains
Budgeting Maximizes single metric Preserves global resource budget
Strategy Generic advice Reward-shaped justification
Memory None RAG memory flywheel (+116% efficiency)

Slide 6 — Memory Flywheel

  • The Numbers: Cold start 42% success rate → Warm (RAG) 88% success rate.
  • The Edge: ChromaDB retrieval lets the agent reason from past successful precedents.

Slide 7 — Current Progress (Status)

  • Live: Flask demo on HuggingFace Spaces.
  • Functionality: 6 working tabs including Comparison, Personality Lab, and What-If Lab.
  • Pipeline: GRPO training backbone complete; model lazy-loads for instant demo reliability.

Slide 8 — Next Steps

  • Full Multi-Step Evaluation: Running 30-day episodes (beyond single-action).
  • Real Data Ingestion: OAuth for Gmail/Calendar signals (currently stubbed).
  • Quantitative Scaling: Benchmarking 1000+ synthetic scenarios.

Demo Script (The 4-Step Sequence)

  1. Stage the Crisis: Open the "Situational Portal". Select Alex (Executive) + Career crisis.
  2. The Cascade: Hit "Start Simulation". Let the 4-frame animation play. Silence for 5 seconds. Then: "Every color change was computed by the graph, zero LLM involvement yet."
  3. The Heatmap: Point at the Red cells. "Red means crisis. Notice how a work deadline dragged Physical Health into the red. The agent must now resolve this composite state."
  4. The Comparison: Switch to "Trained vs Untrained". Hit "Run Comparison". "On the left is the raw model. On the right is the model after RL feedback on our 7-day reward signal."

Counter-Questions & Defensive Positioning (QA)

Question Winning Answer
"Is this just prompt engineering?" "No. We modified model weights via GRPO. The reward comes from the environment simulator, not a system prompt."
"Your environment is hand-coded?" "The environment physics are expert-coded (research-based); the policy navigating them is learned. Chess rules are coded, but AlphaZero is a research breakthrough."
"How do you prevent reward hacking?" "Triple-check: Reasoning audit, resource preservation costs, and discounted 7-day rollouts penalize short-sighted wins."
"Why 1.5B parameters?" "Intentional. It allows consumer-local deployment (privacy) and makes the RL training signal highly measurable."

The Perfect Hook

Opening (30 Seconds)

"Most AI tools give you advice. LifeStack gives you consequences. We built a 6-domain, 23-metric RL environment where a career crisis cascades into sleep loss, relationship strain, and financial pressure—all causally linked. Then we trained a model to navigate that using GRPO. The question we're answering is: can a 1.5B model, trained on life-state rewards, make better long-term decisions than an untrained LLM? We can show you the delta right now."

Closing (The Final Word)

"The real contribution isn't the UI—its the environment + training loop. Everything you see in the demo is an artifact of that system working."