researchengineer.ing
← notebook

February 26, 2026

How and what I am trying to achieve

Note to self: the 300-day objective, the daily shippable artifact protocol, and a hard audit of whether the current setup is sufficient.

I am not trying to "learn more ML." I am engineering a capability shift.

From February 26, 2026 to December 23, 2026, I am running a self-designed, self-enforced program to become an independent research engineer who can:

  • read serious ML/AI papers and reconstruct them from first principles,
  • implement them end-to-end,
  • evaluate and scale them under real constraints,
  • reason mathematically and explain clearly,
  • ship consistently without external pressure.

This is identity-first. I am building the identity of the person who ships.

Program Window (No Ambiguity)

  • Start date: 2026-02-26
  • End date: 2026-12-23
  • Delta between dates: 300 days
  • If I count both endpoints as active workdays: 301 calendar check-ins

The point is daily compounding, not semantic debates about counting.

Daily Artifact System (Ship Every Day)

Every day ends with one committed artifact. Small is fine. Vague is not.

What counts as a daily artifact

A daily artifact is any one of:

  1. Implementation shard
    A function/module implementing one core idea (with minimal test or sanity check).

  2. Experiment shard
    One hypothesis + one run + one result plot/table + one interpretation.

  3. Math shard
    A derivation/proof of a core step from a paper or algorithm (clean, checkable, typed).

  4. Paper reconstruction shard
    A concise reconstruction of one section (assumptions, equations, algorithm, failure modes).

  5. Systems shard
    One profiling/optimization artifact (latency, memory, throughput, bottleneck diagnosis).

Shippable definition (pass/fail)

It is only "shipped" if all are true:

  1. Committed today in git with a clear message.
  2. Concrete output exists (code, plot, derivation, benchmark, or written reconstruction).
  3. Verification exists (test output, numerical check, reproducible command, or logical check).
  4. Insight exists (what I learned, what failed, what I will do next).
  5. Publicly loggable in content/builds/ or content/notebook/.

If any one is missing, it is not shipped.

Daily protocol (default)

  1. Plan (10 min): define today's smallest meaningful artifact.
  2. Deep work block (90-180 min): build/derive/run.
  3. Evidence capture (20-30 min): save output, plot, or checks.
  4. Write-up (20-30 min): concise explanation and failure notes.
  5. Commit + publish (10 min): commit artifact before end of day.

Daily Logging and Review

End-of-day log (mandatory)

For each day, log:

  • Date
  • Artifact slug
  • Type (implementation/experiment/math/paper/systems)
  • Muscle trained (implementation depth, math rigor, experiment design, etc.)
  • Evidence link (file path, command, output)
  • 1 key insight
  • 1 blocker
  • Next-day first action

Weekly review (every 7th day)

  1. Count shipped days.
  2. Count artifact types (am I over-indexed on reading vs building?).
  3. Pick one artifact to extend into a stronger build.
  4. Set next week's default artifact queue (3 minimum).

Monthly integration review

  1. Select best artifacts across the month.
  2. Merge into one stronger narrative build (paper reproduction, benchmark, or system).
  3. Identify missing capability (math, systems, evaluation, writing) and correct allocation.

Setup Check: Is Current Repo Enough?

Short answer: good foundation, not yet sufficient for an uncompromising 300-day system.

What already helps

  • content/builds/_template.md supports rigorous artifacts (theory, implementation, experiments, failure log).
  • make new-build and make quick-note reduce startup friction.
  • Notebook + builds + roadmaps structure supports public shipping.
  • /nudge and /stats give basic progress visibility.

Gaps that will hurt consistency if ignored

  1. Tracking mismatch
    /nudge currently measures only builds, while the UI language implies notebook/external artifacts can count.

  2. Date mismatch with this commitment
    JOURNEY_START is set to 2026-02-20, but this program starts on 2026-02-26.

  3. No explicit daily quality gate in repo
    There is no enforced pass/fail definition for "shippable artifact."

  4. No single daily ledger file
    Without one source of truth, consistency becomes narrative, not data.

  5. Doc drift
    README.md references content/builds/example-build.md, which does not exist.

Verdict

Current setup can support the mission if I enforce the daily artifact protocol above and close the tracking gaps quickly.
Without that, this becomes reading-and-writing theater.

Non-Negotiables

  1. No zero-artifact days.
  2. No private-only progress.
  3. No "I learned a lot" without evidence.
  4. No week ends without at least one artifact worth showing a skeptical reviewer.
  5. No excuses dressed up as standards.

This is not about intensity. This is about consistency with proof.