Skip to content
Read · Build · Ship

A wiki that nudges you towards frontier AI
by Minimum Valuable Builds.

Every pivotal page closes with an MVB — direction, not a copy-paste tutorial. Real HuggingFace IDs, a dataset, a success metric you have to hit, and the structural hints to wire them together. You implement it; the building is where the concept becomes yours.

Sample page → diffusion models Browse the 10 tracks How it's built
71 Substantive concept pages so far
10 Canonical AI tracks (pracha.me)
arXiv Primary sources only — verified at write-time
3 Personas per MVB · applied AI · research engineer · researcher
The compounding journey

Curriculum → Arcs → MVBs

Wikipedia gives knowledge but no journey. roadmap.sh gives a path but no compounding. paperswithcode gives implementations but no pedagogy. This wiki is built around the one thing none of them offer: learning that actually compounds.

1 · Range

Curriculum

One page per concept across the 10 canonical AI tracks. Context that carries forward — by page 3 the reader doesn't re-read what they already know. Reference voice, primary sources only.

→
2 · Depth

Arcs

Opinionated paths from a starting concept to a frontier capability. Curated readings + a compounding-trajectory table. Each step has a named artifact that the next step actually loads.

→
3 · Proof

MVBs

One build per arc step. Persona-tagged — a CS student, a production engineer, and a researcher get different recipes from the same concept. The final artifact is the capstone you can show.

The bet: not tutorial-chasing, not bookmark-stacking, not getting lost in piles of information. A wiki that closes every page with a directed nudge — what to build next, for the persona you are.
The system itself is the work

A closed-loop agentic wiki

Every page is written by a LangGraph pipeline, reviewed by a panel of 8 critics, and held to a checklist the planner built from primary sources. After each cycle a retrospective agent reflects on what worked and what regressed, then seeds the next sprint with the gaps it identified.

1 · Sense

Observer

Scans runs, coverage, citation health, unresolved wikilinks. Computes error signals against quality / coverage / staleness set-points.

2 · Decide

Supervisor

Reads the observation + last retro and rewrites the sprint queue — which topics to generate next, which to revisit, which arc to materialise.

3 · Act

Sprint

N pages built in parallel: research → checklist → write → review (8 critics + knockout selector) → commit. Hallucination guards run on every draft.

4 · Reflect

Retrospective

Scrum-style retro: went-well, went-wrong, needs-depth, what-to-add. Safe items auto-apply (stub-seeds). See the latest backlog.

Built-in hallucination guards. Every draft passes a deterministic gate before review — future-dated arxiv IDs are rejected, pre-verified HuggingFace model IDs are mandatory, the writer's checklist is enforced. The 8-critic panel is the second gate; the knockout selector ensures revisions only land when they improve the page. Full architecture →
Maintained by Prabakaran Chandran · Sources: arXiv · .edu · HuggingFace · GitHub

Home

The 10 canonical tracks

FAIRE is organized around the 10 parallel learning tracks from pracha.me/curriculum. Each subject has its own arcs, concepts, key authors, and builds — slot in at any depth.

Every page is one of four artifact types

Each subject converges on the same shape — concepts, authors, arcs, and builds — designed to feed into each other rather than sit as a bookmark pile.

  • Concepts are encyclopedic, self-contained walk-throughs (Olah/Distill grade), not bullet-point summaries.
  • Authors anchor the field to the people whose work shaped it.
  • Arcs are roadmaps.sh-style learning paths through the concepts.
  • Builds are Minimum Valuable Build recipes — runnable, persona-tagged, real artifact at the end.

Built for three personas

Every page that carries an MVB targets these three. The schema enforces it — any other persona tag is a writing error.

Persona Comes to do Time Their MVB shape
Applied AI/ML engineer (forward-deployed) Ship into production by Friday Half a day – 1 working day Fine-tune a real model and serve it with a measured latency target
Research engineer Reproduce a paper's number on commodity hardware 1–3 working days A reproduced table or figure within ±5% of the published number
Applied researcher Test one hypothesis with one falsifier 2 days – 1 week A 2–3 condition ablation with a plot and a falsification criterion

Every MVB clears the 5-gate quality bar: a real ship-able artifact · a concrete time-to-ship · real HuggingFace model + dataset IDs · a specific success metric · hardness in the middle (fine-tune OR reproduce OR ablate OR deploy — never just pip install + pipeline()).

Source discipline

Every link traces back to a primary source:

  • arxiv.org — papers and preprints
  • *.edu — university lecture notes, course pages
  • huggingface.co — model cards, datasets
  • Official library docs (PyTorch, JAX, Diffusers...)
  • "In production" sections only: official engineering blogs from frontier labs

No Medium. No Towards Data Science. No Wikipedia as a citation.