Delta Brief

Inspiration

EMBA students are usually part-time students and full-time workers. Unlike traditional students, they’re constantly asking: “How does this week’s lesson help me make better decisions at work?” I built Delta Brief to create a positive feedback loop where weekly work realities sharpen in-class learning and discussion, and applied learnings from EMBA classes convert into tangible action items for the workplace.

What it does

Delta Brief generates a one-page pre-class executive brief for the next EMBA session:

• Takes a 60-second weekly check-in (“what changed at work since last class?”)
• Produces 3 ranked moves that matter, tied to the upcoming lesson’s frameworks + learning objectives
• Personalizes using the student’s role, constraints, and capstone goals
• Explicitly updates prior decisions via: Previous → Now → Updated
• Includes class discussion ammo plus a concrete next action + deliverable

How I built it

• Syllabus as source of truth: each session includes a topic, learning objectives, frameworks (name + bullets), prompts, and an assignment hook—so the brief is academically grounded, not generic.

• MemMachine for persistent memory: I integrated MemMachine as the memory layer to store and retrieve student context across sessions.

  • Semantic memory stores stable profile context (role, constraints, capstone).
  • Episodic memory stores weekly check-ins and generated briefs, anchored by session_id = class date. This enables true compounding: Week B can explicitly reference and update Week A’s open threads.

• Strict brief template: forces structure, keeps it to one page, and makes outputs scannable.

• Compare view: a Week A → Week B comparison that highlights what changed and which threads were updated.

Challenges

• Class-light outputs early on: solved by enriching syllabus context and enforcing framework usage.
• Timeline confusion: briefs drifting from syllabus dates; fixed by making session dates canonical everywhere.
• Brittle parsing: relying on markdown was fragile; mitigated by storing structured fields (moves, highlights, thread updates) alongside the markdown.
• Memory retrieval tuning: MemMachine filtering and response mapping required careful handling to reliably fetch the correct prior context.

Accomplishments

• Compounding is visible, not implied: Week B clearly updates Week A through explicit thread resolution.
• Persistent memory that works: MemMachine enables the system to remember prior decisions and constraints across weeks.
• Tight and usable workflow: minimal input, one-page output, and an actionable next step.
• Demo stability: deterministic constraints (framework gating + no repeats + validation) prevent drift during live runs.

What I learned

• Personalization only matters when it’s obviously grounded in constraints and context the student actually cares about.
• Compounding requires explicit mechanics—without them, outputs “reset” each week.
• Episodic memory is powerful, but only if it’s structured and time-anchored; tying memories to session dates made timeline reasoning reliable.

What’s next for Delta Brief

• Extend to semester-scale usage: smarter retrieval, decay of stale context, and better long-horizon continuity.
• Add a “ready for class” automated flow: generate each brief as a PDF artifact and deliver it automatically (via email or text).
• Support multiple course scenarios and richer capstone deliverables.
• Build an evaluation flywheel: track which recommendations led to action, and feed that signal back into memory to improve future briefs.
• Graph-powered decision and stakeholder navigator: By integrating with organizational level data (whether its through the workplace or through the EMBA program), Delta Brief will map weekly work deltas to open threads, decision rights, and dependencies—then recommend the shortest unblock path (who to brief, what artifact to produce), with an explainable decision lineage that feeds the capstone.

Built With

• memmachine
• neo4j
• openai
• postgresql
• react
• typescript
• vite