Who made this

Why this exists

My name is Priyrajsinh Parmar. Praxia started as a personal document. I was navigating the Data and AI landscape without a reliable guide to what to learn, in what order, and why the order matters. The resources existed. The sequencing did not. Most curricula were either too shallow (a list of trending tools) or too academic (a PhD programme without the research context that makes a PhD programme coherent).

What I needed was an opinionated answer to: if you are starting from scratch and want to reach genuine competence across this stack — analysis, machine learning, ML engineering, AI systems — what do you learn, from which sources, in which order, and how do you know when you are done? That question does not have a Wikipedia page. It has, instead, a lot of conflicting blog posts and a lot of people who learned things in the wrong order and are now teaching that order to others.

The map is my answer to that question. It is opinionated because the alternatives are not neutral — they are just less explicitly opinionated. It is curated because comprehensiveness is a trap: a list of 400 resources is not a curriculum, it is a to-do list that will outlive your motivation. I am still walking this path myself. That is not a disclaimer — it is the reason the curation stays honest.

Proof of work

The claim that this map reflects real practitioner knowledge — not curated theory — needs proof. The proof is the work, not the biography.

The flagship project on the Data Analyst stage page — the 15–20 page rigorous statistical report — is described from direct experience. The specification (five research questions, full assumption checks, APA reporting, effect sizes, Benjamini-Hochberg correction, multiple regression with residual diagnostics, written as if for an applied statistics venue) is not aspirational. It is what the project actually required to be done correctly.

Beyond Stage 1, applied projects are on GitHub and reflect the same progression the map describes:

• Conformal prediction on UCI Heart Disease — coverage-guaranteed uncertainty quantification with XGBoost + MAPIE; the Stage 2 UQ project pattern.
• Causal ML / heterogeneous treatment effects — Double ML with SHAP and a deployed dashboard; Stage 2 causal reasoning.
• Real-time ML drift monitoring — PSI, KS test, Evidently in a live monitoring pipeline; Stage 3 production observability.
• Predictive maintenance (before-it-breaks) — LSTM-based Remaining Useful Life forecasting on NASA CMAPSS data; Stage 3 end-to-end time-series.
• Hybrid LLM jailbreak detector — combining multiple signals for 99.88% detection accuracy; Stage 4 applied AI safety.

The exit criteria on each stage are calibrated to what the next stage genuinely assumes. The mathematics sections describe the minimum and research-grade levels because I learned the difference between knowing a formula and understanding its derivation the hard way — by getting stuck on material that assumed understanding I did not have.

The research track is documented from the outside of active research — I have read papers, reproduced results, and worked alongside researchers, but I am not a research scientist. That boundary is marked clearly on the research page: the track describes what the path looks like from the perspective of someone who has studied it carefully and walked the early rungs, not someone who has reached the top.

Editorial voice

Praxia is written in a single voice because it reflects a single point of view. That means it is sometimes wrong. Where the map says “this is the right resource” or “this is the important concept,” that is an editorial judgment — not a consensus of the field, not an optimised algorithm, not a committee decision.

The editorial stance is: honest over encouraging. The time estimates are not aspirational — they reflect what serious, consistent work actually takes. The exit criteria are not low bars — they reflect what you actually need to know to proceed without struggling. The difficulty of Stage 2 mathematics is not softened because many people find it hard. It is hard. Knowing that it is hard, and why, is more useful than being told it is manageable.

Where I am uncertain, I say so. Where a resource may go stale (AI tooling in particular), it is flagged with needs review in the underlying data. Where I recommend something paid, I say it is paid and give the best free alternative.

Contact and feedback

The map is a living document. Resources age; the field moves; the AI Engineer stage in particular will require updates as the tooling stabilises. The footer carries a “links verified as of” date and an invitation to report broken links. Use it.

Substantive feedback — a resource that deserves to be here and is not, a concept that is described incorrectly, a stage that is mis-sequenced — is welcome via the GitHub issue tracker or by email at priyrajsinh03@gmail.com. I cannot respond to every suggestion, but I read them, and the map has already changed because of them.

What is not in scope: requests to add resources because they are popular, comprehensive, or free. The curation criterion is highest-ROI for someone on that specific stage. Popularity is a weak proxy for that; comprehensiveness is sometimes its opposite.