Goals

ML/DL research × systems engineering → quantum computing systems → Summer 2028 research-engineering internship, and entry into quantum-systems research with an artifact behind it

Become a research engineer who can frame an ML/NLP question, build the system, run the experiment, measure the bottleneck, and communicate the result — and who can then carry the same discipline onto a machine that does not work like the one underneath it: state that cannot be copied, gates that are unitary, errors that must be corrected rather than retried. The evidence is sustained lab work, serious research artifacts, and a credible ML-systems portfolio — not course completion alone.

Go deep on two equal pillars and let both bend toward a third. Every day: one paper, one systems chapter, one AI chapter, one system implemented, one open-source contribution. Turn all of it into visible research evidence. Keep competitions deliberately secondary.

The daily contract

Paper: One paper, named in the calendar. Systems or quantum on odd days, AI or an AI-adjacent field on even ones. Done when: A five-bullet card: claim, method, evidence, weakness, next experiment.
Systems chapter: One chapter or section, in the book's own order — OSTEP, CS:APP, DDIA, architecture, parallel, compilers, then quantum information, error correction and quantum systems. Done when: A trace, a test, a design note, a profiler capture — or it fed the day's system.
AI chapter: One chapter or section — the DL/CV/NLP core, then the fields that consume it, then quantum mechanics as physics. Done when: A derivation worked, an implementation, or it fed the day's system.
System: One system implemented. Either whole in a day, or one increment on a long-running project — the day's system is the group-commit path, not 'a database'. Done when: It runs, and one number was measured about it.
Open source: One contribution to someone else's repository. What to contribute is chosen on the day, not scheduled here. Done when: Logged from /Admin — repo, kind, link, what it did, what came of it.

Priorities

Course performance — Protect grades and turn CSE 446/447/451/452 work into artifacts instead of duplicating it with unrelated side projects.
Research capability — ML/DL study and systems study count as research. Build, test, measure, and write; maintain a mentor relationship.
The daily contract — Five lanes a day. It outranks everything below it because it is what compounds — a quarter of kept days is a portfolio, and a quarter of skipped ones is a plan nobody can tell you followed.
Open-source contribution — Something every day; a pull request opened two or three times a week. Log what happened, including nothing — a log with gaps is evidence and a log with invented entries is worse than no log.
Startup — Only continue a cycle when it shares data, infrastructure, or a research question with the main track.
Codeforces — Practice and selected contests for sharpness; never chase rating at the expense of research evidence.
Putnam — Optional proof practice. Pause first under a heavy course, research, or application deadline.

Horizons

Summer 2026 · Aug 13 – Oct 8 — The daily contract actually kept for 57 days. A clean trainer with a gradient check and a deliberate tiny-set overfit; a tokenizer, causal mask and small decoder-only LM with tests; ferrodb crash-recoverable and benchmarked; 57 paper cards; two polished repositories, and 6–10 tailored UW contacts.
Autumn 2026 · evidence and lab entry — One reproducible CSE 446 ML artifact with fixed seeds, a result table and one ablation. One CSE 452 systems benchmark with a traced failure or throughput case. Eight tailored contacts and three substantive conversations, leaving with a trial task, a recurring meeting, or a defined next step.
Winter 2027 · own a research component — A reproduction or bounded extension with fair baselines, multiple seeds and error analysis. Own one lab component — evaluation, data, baseline, reproduction or analysis — and send a concise update every active research week.
Spring 2027 · systems as a research asset — One CSE 451-driven OS, concurrency or storage component with tests, a benchmark, an invariant and a bottleneck explanation. tensorforge differentiating something real. The first quantum papers read — Deutsch, Shor, Grover — while the systems work stays classical.
Summer 2027 · the flagship — Exactly one thesis — efficient transformer inference, LLM evaluation infrastructure, or a small-model training stack with a measured bottleneck. A reproducible evidence package and a submission-quality draft or mentor-backed next decision. The AI reading widens: structural biology, robotics, scientific ML.
Autumn 2027 · ownership, applications, and the quantum turn — A named subproject with stable results, repository and limitations. Resume, project narratives, GitHub landing page and technical report done; at least 12 tailored applications as openings permit. Nielsen & Chuang and Griffiths become the daily chapters, and qsim starts — because the flagship already exists and the applications are already written.
Winter 2028 · publication and interviews — The primary artifact submitted or revised with reproducible code and honest limitations. One measured CUDA/Triton, serving or distributed contribution tied to model evidence. A rehearsed research talk. qroute mapping a circuit onto real hardware connectivity.
Spring–Summer 2028 · close the loop — A handoff-quality stack another collaborator can run from one command path, and an internship executed with one owned contribution a teammate can build on. qdec decoding a surface code, benchmarked against a real decoder, with the threshold plotted and the losses stated.

Pillars

ML/DL → NLP

Six stages of mastery, in order: differentiable programs, learning dynamics, representations and sequence models, NLP foundations, transformers and LM behavior, research-grade evaluation.

Evidence: Each stage ends in code or a result — a numerical gradient check, a repaired overfit, a controlled ablation, a dataset card, a decoder-only transformer with shape tests, a reproduction with a written protocol.

AI beyond the benchmark

After the DL/CV/NLP core, the fields that consume it: structural biology and protein models, robotics and control, scientific computing and neural operators, and the neural-network-as-wavefunction line that connects back to quantum.

Evidence: One reproduction outside the language-model default per period, with a domain baseline you did not invent.

Systems

OSTEP, CS:APP and DDIA first; networks, distributed systems, architecture, PMPP and compilers after; read in the book's own order and never faster than the build.

Evidence: Every chapter ends in a trace, a program, a profiler capture, a design note, or a measured experiment. Never a chapter read for its own sake.

Quantum computing

Quantum mechanics as physics first (Griffiths, then Shankar and Sakurai), quantum information second (Nielsen & Chuang, Preskill, Kaye–Laflamme–Mosca), quantum systems third (compilers, error correction, control, simulation). Entered in Autumn 2027, after the classical systems pillar has shipped.

Evidence: Simulators and compilers written and benchmarked: a statevector simulator, a stabilizer simulator, a qubit router, a surface-code decoder measured against minimum-weight perfect matching. Not notebooks calling Qiskit.

Research practice

One paper a day, named in advance, alternating lanes. One read deeply each week; the rest are structured first passes.

Evidence: A paper card: question, claim, mechanism, strongest baseline, evidence, limitation or confounder, connection to the active project, next action.

Portfolio

Every repository answers one narrow question. The independent tracks ladder ferrodb → riscv-vm → tensorforge → inferd → qsim → stab → qroute → qdec → capstone, and the quantum ones reuse the classical ones rather than starting over.

Evidence: Setup, a run command, tests or checks, a result or benchmark, known limitations, and a statement of personal contribution. By 2028: four or five serious repositories, not tutorial fragments.

How it gets worked

During term: 25–30 hours of course work plus 40–45 extracurricular: The extracurricular budget is 45 hours a week and the floor is 40 — about 6.4 hours a day across seven days, which is exactly what the daily contract costs. That is the source document's original figure, restored.
Where the hours were revised, and revised back: A previous revision cut the term week to 45 hours in total, leaving 15–20 extracurricular, and that figure is what justified pausing the independent systems track inside a quarter. The budget is now 40–45 extracurricular on top of coursework, so the pause is repealed: all five lanes run during term, and the independent track advances every day of the year.
Free weeks: 70 hours: About 7 hours a day — paper 0:45, systems chapter 1:15, AI chapter 1:15, the day's system 2:30, open source 0:45, and half an hour to write down what shipped and what slipped.
The 40-hour floor version: When the coursework week is genuinely brutal: the day's system drops to 1:45 and the log folds into it. The paper and both chapters do not move — they are the lanes that compound, and the ladder desynchronises if one of them skips.
Summer 2027: 50–65 hours a week on one primary role: A research internship, full-time lab, AI2 or adjacent research, or an ML-systems internship. Summer 2028 is the internship itself at 40–55+ hours, with only minimal personal work.
Parallel tracks, sequential gates: Sequential: interface and invariant → correctness → test → benchmark. Data and tokenizer → model → evaluation. Paper reading → targeted outreach. The five daily lanes run in parallel by construction; nothing else should.
The day contract: A day is kept when all five lanes produced their artifact. A day is not kept because the hours were spent. Write which it was.
Every sixth day is catch-up: No new chapter, no new system, no new increment — one paper, and whatever slipped. 124 of the 750 days — every sixth, except the last day of the plan, which should not be a rest day. That is the slack that keeps the dates meaningful instead of turning a bad week into silent abandonment.
Cut in this order: Putnam → extra Codeforces → startup breadth → optional exploration → the day's system shrinks to atomic. Never cut an active mentor commitment, a core course deliverable, or two consecutive days of the same lane.

The minimum viable week

Research: One question, hypothesis or result update for a mentor, a lab peer, or your own ledger.
Portfolio: One improvement: a test, a figure, a result table, a README, a technical note, or a reproducible command.
Open source: Two or three pull requests opened, on top of the daily contribution.
Administration: One outreach, scheduling, application or paper-triage action.
Competition: Only after the above: one Codeforces practice or contest. Putnam only if capacity remains.

What this is not

Do not add every paradigm as another core track. A branch earns a quarter only when you can write the one-sentence research question it changes, the baseline it must beat, and the evidence that would falsify its value.
Do not read everything. A chapter is done only when it leaves code or a result.
Do not start quantum before the classical systems pillar has shipped one artifact someone else has run. It is the third destination, not a third parallel track from day one.
Do not read quantum mechanics as popular science. A chapter counts when the problem set is worked or the simulator reproduces the number.
Do not let the open-source log become a streak to defend. Record the gap; a fake entry costs more than a missed day.
Do not jump to RLHF or PPO before a real research question needs it.
Do not publish shallow updates. Prefer one careful report, reproduction note or engineering retrospective to many.
Do not chase Codeforces rating at the expense of research evidence, and never let Putnam outrank the pillars.
Do not send outreach before understanding the paper and the lab; do not benchmark before correctness or evaluate before a baseline.

Done when

Someone else can see the question, run it, inspect the result, understand the limitation, and identify what you personally owned.
Four or five serious repositories exist — not tutorial fragments — plus short technical reports and a private accomplishment ledger.
A mentor relationship is sustained, not a single conversation: a trial task, a recurring meeting, or a defined next step.
One measured ML-systems contribution exists, tied to model or evaluation evidence.
One quantum-systems artifact exists — a simulator, a compiler pass, or a decoder — benchmarked against an established tool, with an honest statement of where it loses.
The daily log covers 750 days and is true: kept days marked kept, missed days marked missed, and the open-source column filled in by hand.
You can narrate the arc: question → implementation → systems behavior → measurement → written conclusion.

What evidence today makes you more useful to a strong research group? A good answer is an experiment, a system that runs, a measurement, a mentor relationship, or a written result.

Source: Two-Year Execution Schedule and Priorities, Aug 2026 · daily-contract and quantum revision, Aug 12 2026