For full-stack developers moving into AI engineering

Every topic between you and AI engineer, on one map.

Engram lays out 709 topics across 29 domains, from Python internals to agent security. Mark what you already know, study the rest with a written lesson for each one, and always see what to learn next.

Create account See how it works Email and password. Your progress stays in your account.

The whole curriculum. One mark per topic, colored for a sample account. Point at a domain to read it

Strong Understood Practiced Learning Weak Not started

engram / dashboard

Thursday, 8 October

Good evening, Sam

Review 6Continue learning

Completed30%214 of 709

In progress23

Weak4

This week6h 40m

Strong 96 Understood 118 Practiced 31 Learning 23 Weak 4 Not started 437

Up next

Ranked by importance and gaps

Reranking

Weak · unlocks 3 RAG topics · interview critical

35 min

Async database access

Used by Enterprise Knowledge Assistant

40 min

Tool calling

Prerequisite for Agent loop

30 min

Agent evaluation

High interview importance · not started

45 min

Needs attention

4 weak

Reranking

RAG · confidence 2

Weak

Redis Streams

Redis / Caching · review due

Weak

Isolation levels

Databases · missed in interview

Weak

Prompt injection

AI Security · confidence 2

Weak

29domains

709topics

709written lessons

303prerequisite links

70interview questions

10portfolio projects

Curriculum

Start from what you already know.

You have shipped software before, so the curriculum doesn't start you at zero. Skim the tree, set a status in one click, and the topics you already know drop out of the way.

Filter by status, confidence, importance, interview relevance, track and practical work.
Edit inline. Change status, set confidence, mark weak or add a note without leaving the list.
Search everything across topics, subtopics, projects and questions.⌘K

engram / curriculum / rag

Software Engineer

Python88%

FastAPI / Backend74%

Databases61%

System Design38%

AI Engineer

Transformers / LLMs44%

Embeddings72%

RAG41%

Agents18%

AI Evaluation9%

Production6%

RAG

24 of 60 complete

Retrieval

StrongDense retrievalEmbed the query, nearest neighboursInterview100%

UnderstoodChunking strategiesFixed, recursive, semantic, by structurePractical86%

LearningHybrid searchBM25 plus vectors, rank fusionInterviewPractical42%

WeakRerankingCross-encoders, top-k, latency budgetInterview18%

Not startedQuery rewritingHyDE, multi-query, decomposition0%

Knowledge map

See how the two halves connect.

The software engineer tree and the AI engineer tree sit side by side, joined by the real prerequisite links between them. Pin a domain to see what it unlocks and what it depends on.

Prerequisites first. A topic whose prerequisites are open gets pushed down your list, and the prerequisite gets pulled up.
Four kinds of link: prerequisite, builds on, related and commonly confused with.

engram / map

Software Engineer7 domains

Python88%

FastAPI / Backend74%

Databases61%

Redis / Caching52%

Queues33%

System Design38%

AI Engineer20 domains

Embeddings72%

Vector Databases46%

RAG41%

Agents18%

Memory10%

Production6%

Learning mode

A written lesson for every topic.

Open the top recommendation and study in one focused view: the lesson, subtopics to tick, a practical task, interview questions and your own notes. The timer logs the session when you stop.

Mark as strong, mark as weak, or review later as you finish, from the keyboard.s · w · r
Set confidence from 1 to 5 with the number keys.1–5

engram / learn / reranking

RAG / Retrieval / Reranking

Reranking

18:42

WeakConfidence 2/5Interview HighUsed in 2 projects

Why a second pass

First-stage search is built for recall: a bi-encoder embeds the query and each document separately, so you can search millions of vectors fast. A cross-encoder reads the query and a passage together and scores how well they match. It is more accurate, but it cannot precompute anything, so you run it only on the top 50 or so candidates.

Treat top-k as two numbers: a large k for recall, then a small k after reranking for the prompt.

Bi-encoder vs cross-encoder

Choosing candidate k and final k

Latency budget and batching

Mark as Strong S Mark as Weak W Review later R

Review and interview practice

Find the weak spots before an interviewer does.

Topics you mark weak come back for review right away. Reveal the answer, rate yourself, and the next review moves out or comes closer. Interview practice works the same way, with answers hidden until you ask.

Eight interview areas: Python, Backend, AI, RAG, Agents, System Design, DSA and Behavioral.
Reveal and rate without touching the mouse.space · 1–4

engram / review

RAG · RerankingDue todayReviewed 3 times

When would you add a cross-encoder reranker, and what does it cost you?

When first-stage recall is fine but the best passages aren't landing at the top.

Run it on roughly the top 50 candidates, then keep a small final k.
Costs tens to hundreds of milliseconds per query, so truncate and batch.

Strongin 8 days Okayin 5 days Weakin 2 days Forgottomorrow

2 of 6 dueNext: Redis Streams

Projects

Prove it with something you built.

Ten portfolio projects, from a first LLM API app to a production AI full-stack application. Each one links to the topics it exercises, so you can see which gaps to close before you start.

Milestones drive progress. Tick them off, add your own, keep notes per project.
Skill readiness shows which linked topics are ready and which are still weak or untouched.

engram / projects / enterprise-knowledge-assistant

Project · Advanced

Enterprise Knowledge Assistant

In progress3 of 8 milestones6 of 9 skills ready

Milestones

38%

Ingest and chunk the docs

pgvector search endpoint

Sign-in and per-team access

Rerank and cite sources

Stream answers to the UI

Skills

3 gaps

FastAPIStrong

pgvectorUnderstood

AuthenticationStrong

RerankingWeak

StreamingLearning

RAG evaluationNot started

How progress works

No streaks, no points. Just an honest model.

You set status and confidence by hand. Everything else is arithmetic you can read.

Not started0%

Learning25%

Practiced55%

Understood80%

Strong100%

Weak counts as 20% and puts the topic in your review queue right away.

topic progress = 70% status + 15% subtopics done + 15% confidence

Completed

Understood or Strong. Domain and area totals weight each topic by its importance.

Weak

Marked weak, or confidence 2 or lower on a topic you've started.

Up next

Open topics ranked by importance, interview relevance, weakness, work in progress and project use.

Reviews

Strong stretches the gap 2.5×, Okay 1.6×, Weak halves it, Forgot resets it to a day.

Keyboard

Fast enough to use every day.

⌘KSearch anything

gcCurriculum

glLearning mode

grReview queue

1–5Set confidence

swrStrong, weak, later

spaceReveal the answer

?All shortcuts

Open the map and mark what you already know.

It takes a few minutes to set your starting point. After that, Engram tells you what to study next.

Create account Sign in