GenPHD: Decision Intelligence for AI Engineers

Turn conflicting AI advice into an evidence-backed next build action — then learn from what happens when you act on it.

Comment

Inspiration

Every AI engineer has lost hours asking the same question in five different tabs — ChatGPT says use agents, a tutorial says use a framework, the docs say something else entirely. The problem was never a lack of answers. It was a lack of a system that remembers your project, weighs evidence over opinion, and turns a decision into an actual next step. That's why we built GenPHD — the decision intelligence layer for AI engineers.

What it does

GenPHD turns conflicting AI advice into an evidence-backed next build action, then learns from what happens when you act on it.

The core loop:

  1. You describe your goal, active project, stack, time budget, and current blocker.
  2. GenPHD builds a concise roadmap with your next three milestones.
  3. When you're stuck, you ask a real decision question — e.g. "Should I use LangGraph for this two-day RAG project?"
  4. GenPHD returns a Decision Brief: source-backed evidence, tradeoffs, a recommendation, an explicit confidence level, a counterfactual ("choose the alternative if…"), and one next action.
  5. That action becomes a Build Mission with a target outcome and acceptance criteria.
  6. Once you complete it, GenPHD records the outcome, updates your skill evidence, and adjusts your roadmap.

It is not a generic chatbot, a course platform, or a multi-agent dashboard. It's a closed decision loop — evidence in, action out, learning compounding over time.

How we built it

  • Frontend/App: Next.js (App Router, TypeScript strict mode), Tailwind CSS, shadcn/ui, Lucide icons
  • Backend: Supabase (PostgreSQL, Auth, Row Level Security, Storage), typed server routes, Zod schema validation at every boundary
  • AI workflow: OpenAI as the primary reasoning layer, orchestrated as a controlled pipeline — Context Builder → Evidence Retriever → Parallel Deliberation → Claim Adjudicator → Action Composer → Reflection Evaluator — not an unconstrained autonomous agent
  • Architecture: Deliberate modular monolith for the MVP; services split only when there's a measured reason (e.g. sandboxed code evaluation)
  • Design system: A monochrome, restraint-first UI (Apple-level calm, Linear-level density discipline) — no gradients, no gamification, one primary action per screen
  • Source grounding: Curated, versioned source corpus with visible URLs and dates, tiered by trust (official docs > maintainer repos > practitioner articles > unsourced social content)

Challenges we ran into

  • Designing a Decision Brief structure that's genuinely useful rather than "AI says X" — we had to force every recommendation to expose uncertainty and a counterfactual, not just a confident-sounding answer.
  • Resisting the urge to build a flashy multi-agent dashboard. The real value was in a disciplined, inspectable workflow, not agent theatre.
  • Keeping memory transparent and user-controlled (visible, editable, deletable) while still making the roadmap feel adaptive and smart.
  • Scoping the MVP tightly — cutting gamification, social feeds, and admin dashboards that would have diluted the core loop.

Accomplishments that we're proud of

  • A working end-to-end loop: onboarding → Decision Brief → Build Mission → reflection → updated roadmap.
  • A Decision Brief that always shows its evidence, its confidence, and what would change the recommendation — instead of presenting model output as ground truth.
  • A design system that stays calm and legible under real product complexity, with zero dashboard clutter.

What we learned

Model consensus isn't trust — evidence quality, recency, and fit to the user's actual constraints are what make a recommendation defensible. Building a product around a loop (decide → act → reflect → improve) creates far more lasting value than another single-turn chat interface.

What's next for GenPHD

  • Repository-aware evidence retrieval tied directly to a user's codebase
  • Secure sandboxed code evaluation for Build Missions
  • Source freshness alerts when a past decision's evidence changes
  • Shared/team decision records for bootcamps and engineering teams
  • Portfolio-ready, source-backed skill evidence

Built With

Share this project:

Updates