12 | Portfolio Architecture¶
“Your portfolio is not a collection of projects — it is a proof ladder for your pitch sentence.”
Your pitch sentence at Month 13 is this:
“I’m an applied ML engineer who rebuilt the discipline from first principles over 13 months — from linear algebra and probability through transformers, diffusion, and LLM alignment — and can now design, implement, debug, and take to production any ML system, read and critique the research behind it, and explain it to anyone; one of a few hundred practitioners globally who are genuinely both builders and scholars.”
Every word in that sentence is a claim. Every rung in this portfolio is the evidence for a specific claim. If you cannot point to a public artifact that substantiates the claim, the sentence is fiction. The portfolio does not exist to impress — it exists to prove.
The 8 Rungs¶
Rung |
Month |
Phase |
What You Build |
Published At |
What It Proves |
|---|---|---|---|---|---|
1 |
2 |
Phase 0: Foundations |
PCA + Gradient Descent from NumPy scratch |
GitHub |
You understand what you’re doing, not just calling library functions |
2 |
4 |
Phase 1: Classical ML |
Full Kaggle pipeline + post-mortem |
Kaggle + Blog |
End-to-end classical ML competence and analytical thinking |
3 |
6 |
Phase 2: Deep Learning |
Neural network from scratch (micrograd/makemore-style) |
GitHub |
You understand backprop, not just |
4 |
8 |
Phase 3: Architectures |
Transformer lab — nanoGPT or ViT fine-tune |
GitHub + HuggingFace + Blog |
You can work with modern architectures and communicate findings |
5 |
10 |
Phase 4: LLMs |
QLoRA fine-tune + RAG system + live deployment |
HuggingFace Spaces + GitHub |
You work with LLMs at production-adjacent depth — not API calls |
6 |
12 |
Phase 5: MLOps |
Complete ML pipeline: versioning → serving → monitoring |
Live system + GitHub |
You are a senior-level ML engineer who ships and maintains |
7 |
13 |
Phase 6: Research |
Independent paper reproduction with analysis |
GitHub + Blog |
You can read, implement, and critique research |
8 |
13 |
Phase 6: Writing |
3–5 technically rigorous blog posts with engaged readership |
Substack/Medium/Blog |
You contribute to the field’s knowledge, not just consume it |
⚠️ Rungs 5 and 6 are HARD GATES. You do not advance to Phase 5 without Rung 5 complete. You do not sit for senior-level interviews without Rung 6 complete. Partial credit does not exist at a gate.
Why These Two Hard Gates?¶
The logic is straightforward and unsentimental. The skills these two rungs test are the ones most frequently faked on resumes and most immediately exposed in technical interviews. LLM engineering (Rung 5) is now baseline expectation at most applied ML roles in 2025 — but calling the OpenAI API is the most common impersonation of real LLM work. Production ML systems (Rung 6) permanently separate researchers who can prototype from engineers who can ship something that doesn’t break at 3 AM. Both gates require live, demonstrable, public artifacts. An interviewer can ask: “Can you show me the endpoint?” If the answer is no, the rung is not complete.
The Skill Progression Arc¶
Each arrow is a phase transition, not a skill increment. You are not adding a tool to your toolkit — you are changing the category of what you can do. Math→Code is the gap between knowing and doing. Code→Systems is the gap between doing and architecting. Systems→Frontier is the gap between architecting and shipping at scale. Frontier→Research is the gap between executing known patterns and contributing original thought.
The Signal Gradient¶
The rungs are calibrated so that each one answers a specific question a hiring manager would ask:
Rung 1 answers: “Do you actually understand ML mathematics?”
Rung 2 answers: “Can you think end-to-end about a real problem?”
Rung 3 answers: “Do you know what’s happening inside the model?”
Rung 4 answers: “Can you navigate the modern transformer ecosystem?”
Rung 5 answers: “Can you work with frontier models beyond API calls?”
Rung 6 answers: “Can you build and maintain a real production system?”
Rung 7 answers: “Can you engage with and implement original research?”
Rung 8 answers: “Can you advance collective understanding in the field?”
A portfolio missing Rungs 1–3 is a senior candidate whose foundations are invisible. A portfolio stopping at Rung 4 is a strong junior. Rungs 5–8 are what the 13-month pitch sentence requires to be non-fiction.
Publication Standards¶
Every rung must be publicly verifiable. No private repos. No “available on request.” No “I did this at work but can’t share it.” The public artifact is the proof — if a hiring manager cannot find it in 90 seconds via your GitHub or portfolio page, it does not count.
Platform |
Used For |
Rungs |
|---|---|---|
GitHub |
All code, READMEs, documentation |
1, 2, 3, 4, 5, 6, 7 |
HuggingFace Hub |
Model cards, Spaces demos |
4, 5 |
Kaggle |
Competition submissions, public notebooks |
2 |
Substack / Medium / Personal Blog |
Technical writing and post-mortems |
2, 4, 7, 8 |
Announcement + professional framing per rung |
All |
Hard Rules for Every Rung¶
Git history is evidence. Single-commit repos look like panic uploads. Incremental commits show actual development process.
README.mdis not optional. First three lines: what it does, why it matters, how to run it. A README-less repo is an abandoned repo.No hardcoded secrets. API keys, tokens, credentials committed to a public repo is an immediate disqualifier — not just for the rung, but for the candidate.
Quantify or don’t claim. “Achieved good results” is meaningless. “Achieved 87.3% F1 on the test set, top 18% on the public leaderboard” is evidence.
Document failures. A “Key Learnings” or “What Didn’t Work” section is one of the highest-signal things you can put in a README. It shows intellectual honesty and genuine engagement with the problem.