I'm a software engineer who took a 6-year break from the field for mental health reasons. When I came back, I found an industry that had been transformed by AI — and a skills gap I needed to close before I could build what I actually wanted to build.
So I built a doctoral-level self-study curriculum. Here's what that actually looks like.
The curriculum
phd-applied-ai is 67 modules across 6 phases:
- Phase 0 — Math foundations: Linear algebra, calculus, probability, statistics, information theory
- Phase 1 — CS fundamentals: Computability, complexity, algorithms, type theory, distributed systems
- Phase 2 — ML foundations: Classical ML through transformers, reinforcement learning, graphical models
- Phase 3 — Advanced AI: LLMs, fine-tuning, diffusion, multi-agent systems, interpretability, safety
- Phase 4 — Specialized research: AI in education, adaptive learning, ethics, affective computing, neurodiversity
- Phase 5 — Thesis and defense
Every module includes: a theory digest at doctoral level, a working code project, a mastery rubric, and a dated study log. The whole thing is CC BY-NC-SA and live on GitHub now.
The par scoring system
Traditional grades and deadlines are incompatible with variable-energy learning. I replaced them with par scores — a borrowing from golf.
Each module has defined par criteria: qualitative standards that describe what "good enough to move on" looks like. You hit par when you hit it. Over par means you went deep. Under par means you moved fast. Neither is a failure. There's no "you're behind" state.
I have ADHD and bipolar disorder. Deadline-based pacing has never matched how my brain actually operates. The par system makes that mismatch structurally irrelevant.
The advisor committee
Five AI advisor personas, each calibrated to a distinct OCEAN personality profile:
- Dr. Chen (ML Theory) — high Conscientiousness, high Openness; demands formal precision
- Dr. Kowalski (Systems/Architecture) — high Agreeableness, low Neuroticism; asks what breaks at scale
- Dr. Williams (Education/Pedagogy) — high Agreeableness, high Openness; asks if this serves the learner
- Dr. Okonkwo (AI Ethics) — high Openness, low Agreeableness; asks whose values are encoded
- Dr. Patel (Applied AI/Industry) — high Extraversion, low Conscientiousness; asks if it ships
Running a module debrief means defending your understanding against five different critical lenses simultaneously. The adversarial diversity surfaces things passive review misses — a high-Conscientiousness reviewer catches methodological gaps that a high-Openness reviewer would skip past.
The personas are model-agnostic prompts in the repo. Drop any PERSONA.md into Claude, GPT-4o, or a local Ollama model as a system prompt.
The research angle
I'll be logging energy, focus, and mood for every study session over the next 5+ years. The AIED research community has almost no longitudinal data on how neurodivergent learners navigate self-directed doctoral education. Mine will be public — released as the curriculum progresses, committed to the repo as session logs.
The curriculum is also the proof-of-concept for Hearth & Code, an adaptive AI-native learning platform I'm building. Every module I study is a prototype of what the platform will eventually offer. The data I collect as a learner is the early research base.
Three design decisions I'd like feedback on
- Module sequencing in Phase 0–1 — I put statistical foundations depth (M02–M04) before gradient-based methods. Does that sequencing look right to people who've studied this material?
- Thesis framing — I'm treating this as design research on a neurodivergent adaptive learning system, not traditional ML benchmarking. Methodological thoughts welcome.
- OCEAN advisor model — any prior art on personality-modeled AI tutors or adversarial committee simulations?
Repo: github.com/hearthandcode/phd-applied-ai
Blog: hearthandcode.substack.com
Top comments (0)