Operating plan · three-to-six-month runway · 22 Aug 2026
Each batch opens on a gate you have met, not on a date. If a gate is not met, the next batch is not bought, and that is information rather than failure. The dates against each batch are the earliest they could plausibly open, never a trigger. Every layer in the table below names the gate that enforces it; a layer with no gate against it will not happen.
Layers are listed in the order they are done and link to their blocks in resource-library.html. Hours assume 15 h/week (3 h × 5 days).
Six STAR stories, written. ~3 h. Four lines each: context in one sentence, the decision you made in first person singular, the number, what you would do differently. Story 2, the ambiguous problem you scoped yourself, decides the Principal deep-dive; confirm it exists before writing the other five.
Opening rehearsed twenty times, recorded. ~2 h. The script exists verbatim; twenty reps is two hours.
Control question recorded: forty accounts, three devices. 2 min. Fluent or fumbled, block 12 is the work; the recording sets which primers you skip.
CV fixes #3 and #4. 15 min. Narrow "Kubernetes" to what you can defend; move the four patents into the summary. Fix #3 replaces 22 hours of Drill 9.1 for every loop whose job description does not require Kubernetes, which is most of them.
First two applications logged. ~2 h. Tier C or tier B. Columns: date, company, role, level, outcome.
Gate: Batch 0 study opens when the six stories exist. About 10 h in total, one week at 15 h/week with the hiring-manager interview inside it. AlgoMonster waits a week.
| # | Layer | Hours | Cost | Enforced by | Why it sits there |
|---|---|---|---|---|---|
| 0 | Applicationstwo per week, logged · from week 0 | 2 h/week~32 h over 16 weeks | Free | Batch 0 gateeight logged in four weeks · Batch 1 study does not open without them | Two a week, logged; one of the two can be a recruiter. Applying unlocks studying: this gate runs the other way round from the purchase gates. |
| 1 | NarrativeSTAR stories and two ADRs · week 0 → Batch 1 | ~14 h6 stories 3 h · ten 5 h more · ADRs 6 h | Free | 00 gate · Batch 1 gatesix stories this week · ten stories and two ADRs by Batch 1 | At your level it decides more offers than SQL does, and the nearest event is a deep-dive, not a screen. Library block 12. |
| 2 | Spoken practicemocks with humans · Batch 0, then weekly | ~4 hthen 1 h/week | Free | Batch 0 gatetwo spoken mocks done | Highest return at zero cost. The skill is talking and typing at the same time, and it is the one thing never trained. |
| 3 | Mechanical recoveryalgorithms and syntax · Batch 0 | ~15 h20 sessions × 45 min | $189 | Batch 0 gate20 logged sessions in 4 weeks · blank-file test re-taken and passed | The one measured failure. Sessions measure attendance and the test measures recovery, so the gate has both. A pattern retires after two clean passes a week apart; without the Sunday maintenance hour it is gone again in six months. |
| 4 | SQL and pandasdiagnostic · Batch 0 | ~1 h+15 h remediation if red | $0–60 | Batch 0 diagnosticsmeasured · free remediation from the day a test fails · paid platform only in Batch 2 | Probably syntax rust rather than judgement, and it may resolve for free, which is why it is measured before anything is bought. The LeetCode study plans start the day a test goes red. Library drill 11.2. |
| 5 | LLM evaluation vocabularyPrime 8 · Batch 0 · book conditional | ~5 hbook +15 h if bought | $0 / ~$45 | Batch 0 gatePrime 8 self-check answered · book only in Batch 2 | The live role is generative, with a proprietary foundation model to evaluate. Your Oracle pre-release evaluation work is the card; Prime 8 (Arize, Evidently, two talks, all free) gives it 2026 vocabulary. Library block 8, primer only; the drills are conditional. |
| 6 | Depth in recsysevaluation, biases, two-stage · Batch 0 → Batch 1 | ~12 h8 h reading · 4 h seed repo | Free | Batch 1 gateone evaluation question answered cold · seed repo exists | It decides the specific offer in front of you and nobody sells it; it turns experience into an argument rather than a background detail. The seed repo: ndcg@k, recall@k, a temporal split. |
| 7 | ML system designspoken, not read · Batch 1 | ~30 hPrime 5 h · five designs 15 h · read-rebuild-diff 10 h | ~$45 | Batch 1 gatefour recorded designs · tier A opens on five written, one delivered verbally | Where senior separates from Principal. One book and practice out loud; no platform reaches this far. Library block 10. |
| 8 | Statistics and ML, rehearsedBatch 1 | ~3 htwo recorded answers | Included | Batch 1 gatetwo answers inside ninety seconds | Your strong point. The risk is answering like an academic when they asked for ninety seconds and a decision. Every DS loop asks this, including the live one. |
| 9 | Real-time serving and latencymonths 3–4 | ~34 hPrime 6 h · 1.1 10 h · 1b 4 h · 1.2 14 h | Free | Batch 3 entry gatelatency table and parity test in the repo | The largest capability gap for target #1 and nothing a platform sells. After the live loop has resolved one way or the other; runs entirely on your M-series. Library block 1. |
| 10 | One public repo, production standardmonths 3–4 | ~24 hPrime 4 h · hardening 20 h | Free | Batch 3 entry gatepublic · strict typing · tests · CI green | The one public artifact. Hardens whichever repo is strongest by then: the recsys-eval seed if VistaPrint is still live, the serving repo if a fraud loop is. One, not both. Library drill 11.1. |
| 11 | Graph methods for fraudconditional | ~18 hPrime 6 h · 3.1 12 h | Free | Conditionalno gate unless a fraud or risk loop is live | Opens on a loop whose job description says rings, mules or multi-accounting. The cross-cluster patent is adjacent; say so in interviews whether or not the drill is done. Library block 3. |
| Eight of the twelve layers cost nothing · ~160 h core, ~190 h with conditionals, plus 2 h/week applying: under 15 h/week across 16 weeks | Eleven of the twelve are gated | ||||
| Batch | Spend | What you buy | What you do with it | Gate |
|---|---|---|---|---|
| Batch 0earliest expected: weeks 1–4 · load ≈ 50 h | $189 | AlgoMonster lifetime Plus everything free: LeetCode study plans, Tech Interview Handbook, the NeetCode roadmap, Exponent mocks, Exponent's system design guide, Udacity's A/B Testing course, and Prime 8's two evaluation guides (Arize, Evidently). | Mechanical rebuild, the five diagnostic tests, Prime 8, the seed repo
A log with a next-review column. Two spoken mocks. The first artifact: ndcg@k, recall@k and a temporal split written by you, in a repo. Two applications a week, logged. |
20 logged sessions in 4 weeks · blank-file test re-taken and passed · two spoken mocks · eight applications loggedPlus the five tests taken and written up, and Prime 8's self-check answered. Fewer than twenty sessions: Batch 1 is not bought. Fewer than eight applications: Batch 1 study does not open either, not Interview Query and not the design Sundays. |
| Batch 1earliest expected: month 2 · load ≈ 55 h | $244–344 | Interview Query + Chip Huyen, Designing ML Systems Annual at $199 if the gate was met with difficulty; lifetime at $299 if you cleared it comfortably, because at that point usage is proven and the bet becomes arithmetic. | Breadth, and Principal level Interview Query by target company, never linearly. From the book, one chapter a week and one 45-minute spoken design, recorded, every Sunday: that Sunday design is the maintenance hour, not an addition to it. Stories 7 to 10 and the two ADRs. Two statistics answers, recorded and timed. | Four recorded designs · ten stories and two ADRs · two statistics answers inside ninety seconds · one recsys evaluation question answered coldListened back to by you. Cold means offline evaluation, popularity bias or two-stage retrieval, out loud, no notes, no preparation on the day. Tier A (Feedzai, Revolut, N26, Nubank, MercadoLibre, Vinted, Datadog) opens on five written designs and one delivered verbally, not on a week number. |
| Batch 2earliest expected: months 3–4 · conditional · load ≈ 60–110 h | $0–105 | Only what the diagnostic marked red DataLemur at about $60/year if SQL failed, or StrataScratch monthly if pandas is the part still stuck, one or the other, never both, under the one-platform-per-layer rule. AI Engineering at about $45 only if a generative role reaches a technical round. | Close measured holes, and build the two free layers no platform sells Every purchase in this batch needs a failed test with a date behind it. Alongside, and free: layer 9 (34 h) and layer 10 (24 h). Layer 11 (18 h) only if a fraud loop is live. | Final rounds confirmedBefore you have finals, Batch 3 makes no sense. The material Batch 3 needs is its own entry gate, below. |
| Batch 3earliest expected: months 5–6 | $180–300 | One calibration session With someone who actually interviews at Principal level. Interviewing.io or IGotAnOffer. | Two or three weeks before the final With material in hand. It is the best signal per euro in the whole plan. | Entry gate — material in handThe repo public and hardened (layer 10), the latency table and parity test inside it (layer 9), the designs written up rather than only recorded, the stories rehearsed out loud. This is the one spend that is wasted if you arrive empty-handed, so its condition sits in front of it rather than behind it. |
| Six-month total | $613–938 | Of which $189 is already decided · the rest opens only if the previous gate is met | ||
Structy, Educative (including Grokking the ML Interview), Deep-ML, Exponent annual, Final Round AI. The first three duplicate what you will already have; the fourth charges you for courses when the part of Exponent you need is free; the fifth is a disqualification risk. This line exists so that in three months, when a promotion appears, the decision is already made.
Permanent rule: one platform per layer. The moment you have two that do the same thing, cancel one. It is why DataLemur and StrataScratch are an either/or in Batch 2 rather than a pair.
The resource library is not a purchase list. It names interviewing.io, DataLemur, StrataScratch and Educative as resources, and its cost appendix prices infrastructure, not platforms. Purchases are decided here, by these gates, and nowhere else. Where the two files disagree, this one wins.
At the end of the six months, what you are left with is not the subscriptions. It is a repo with ranking metrics you implemented yourself and a temporal split with no leakage, public and hardened. Four to six system designs recorded and then written up. Ten STAR stories with numbers in them and two ADRs. A log of which patterns you fail and how often. An application log with outcomes. That survives any expiry date and works again in the next cycle.
The repo and the designs are publishable. An article on offline evaluation in recommendation, with code, is worth more than a certificate, and at fifty-four recent visible output outweighs catalogue consumed.
Maintenance between cycles. Two hours every Sunday, indefinitely: one of spaced repetition on algorithms, one of spoken design. Without it, everything above depreciates.