SELF-PACED AI INTERNSHIPS

Do the work an AI engineer does.

Every internship builds a system companies ship today, grounded in the papers behind it — with AskThili, an AI mentor, beside you the whole way.

No card to start · 175 ⚡ free — enough for your first beginner or intermediate internship, end to end

The techniques you'll build are running in production at

Google·OpenAI·Anthropic·NVIDIA·Meta·Hugging Face·Microsoft·GitHub·Netflix·Uber·Amazon·Spotify·AWS·Cursor

Named for where a technique is used. No affiliation or endorsement implied.

Real
Real projects industry ships.

Not toy datasets. An agent memory engine, a DDoS detector, a dynamic-pricing engine — systems in production at named companies right now.

Grounded
Backed by the papers.

Every internship traces to the research that invented the technique — the landmark papers themselves, linked and read, not summarised.

Mentored
A mentor who never leaves.

AskThili knows the exact section you're stuck on. Ask at 2am and get an answer about your problem — not a generic search result.

How an internship runs.

Four steps, in order. You read the build, then you actually build it — the same way the job works.

STEP 1Enroll free

Make an account and enroll. Your 175 ⚡ starting balance covers your first internship — no card, no subscription.

STEP 2Read the build

The complete notebook, every cell and its real output, rendered in your browser. Nothing to install to start reading.

STEP 3Build it yourself

Then write it in your own environment, the way you would at work — with AskThili answering from the section you're on.

STEP 4Submit and get scored

Push your repo and take the completion test. You leave with a working system and a score you can defend.

One internship. Not a tutorial.

Every one is built three ways — so you finish with a system, not a viewing history.

AI SDLC · Advanced · 9 sections · ~1h 12m · 200
🛠 Parse a Python repo into a navigable call/import graph using only the stdlib ast moduleWhat you'll do
The code-intelligence layer behind Sourcegraph, GitHub code search, and the repo-level retrieval every serious coding agent needs before it can safely edit an unfamiliar codebase.Where it's used
📄 From 2 original research papersGrounded in research
💬 Stuck? Ask AskThili — it knows this exact section.Mentor, in context

Browse every internship. No account needed.

See exactly what you'd build, where it's used, and which papers it's built on — before you sign up for anything.

🧭 Or browse by goal — curated learning paths →
founder

Agent Blast Radius

Advanced· ~40 min
🛠 Tell an agent that answers apart from one that acts, and explain why "automation" hides that difference
The scope-review conversation that happens right after "let's automate this" and right before an agent gets write access to something real.
200
localbiz

Automated Money Management

Beginner· ~40 min
🛠 Extract structured data from a receipt photo and know when to double-check the extraction
The moment right after CoWork for Your Shop, when "can it help with my finances too" is the obvious next question.
50
AI SDLC

Build a Codebase Knowledge Graph

Advanced· ~1h 12m
🛠 Parse a Python repo into a navigable call/import graph using only the stdlib ast module
The code-intelligence layer behind Sourcegraph, GitHub code search, and the repo-level retrieval every serious coding agent needs before it can safely edit an unfamiliar codebase.
SourcegraphGitHub
📄 2 papers200
AI SDLC

Build a Disciplined Coding Agent

Advanced· ~1h 20m
🛠 Build a plan-gate and a critique-gate for a coding agent, and wire them into one toggleable loop
The discipline behind Claude Code's plan mode, Cognition's Devin, Cursor, and GitHub Copilot Workspace — all publicly wrestling with the same reliability problem this course builds from scratch.
AnthropicCognitionCursor
📄 2 papers200
MLOps & tooling

Build a Model-Serving Engine

Advanced· ~1h 20m
🛠 Build a model-serving engine from scratch — the prediction API, batching queue, and cache behind a production inference service
Every company serving ML in production runs this layer — Netflix's recommendation services, Uber's Michelangelo, NVIDIA Triton, and Ray Serve are all the pieces you build here, industrialized.
NetflixUberNVIDIA
📄 2 papers200
Agents & LLM apps

Build an Agent Memory Engine

Intermediate· ~1h 20m
🛠 Build long-term memory for AI agents
Behind the "memory" features in ChatGPT and Claude; personalized assistants and tutors; customer-support bots that recall past tickets; and long-running autonomous agents — productized by memory stacks such as Maximem Synap, Mem0, and Zep.
MaximemMem0Zep
📄 3 papers150
Agents & LLM apps

Build a ReAct Agent

Intermediate· ~1h 12m
🛠 Build the ReAct loop from scratch in pure Python — reason, act, observe, repeat
The core loop inside every tool-using agent — OpenAI function-calling agents, Anthropic's Claude tool use, LangChain and LlamaIndex agents, Perplexity's answer engine, and AutoGPT-style task runners.
OpenAIAnthropicLangChain
📄 1 paper150
recsys

Build a Recommender from Scratch

Advanced· ~1h 44m
🛠 Build a recommender from scratch — similarity, collaborative filtering, and matrix factorization over a real ratings dataset
Amazon's item-to-item recommendations, the Netflix Prize solutions, and Spotify's early collaborative filtering all run on the techniques you build here; every modern recommender is measured against exactly these baselines.
NetflixAmazonSpotify
📄 5 papers200
recsys

Build a Two-Tower Retrieval Engine

Advanced· ~1h 28m
🛠 Build a two-tower retrieval model — separate user and item encoders trained into one shared embedding space
YouTube, Netflix, Spotify, and Pinterest all run two-tower retrieval in front of their rankers; it is also the architecture behind modern semantic search and RAG retrieval.
NetflixYouTubeSpotify
📄 4 papers200
Language models

Build GPT from Scratch

Advanced· ~1h 20m
🛠 Build a decoder-only GPT language model from scratch in PyTorch and generate text with it
The core architecture behind every modern LLM — OpenAI's GPT, Google's Gemini, Anthropic's Claude, and Meta's Llama all run decoder-only transformers like the one you build here.
OpenAIGoogle DeepMindAnthropic
📄 2 papers200
Agents & LLM apps

Building a Coding Agent

Advanced· ~1h 44m
🛠 Build a coding agent that reads, edits, and runs code in a tool-use loop
The engine inside AI coding assistants — Claude Code, Cursor, Aider, GitHub Copilot Workspace, Devin — an LLM in a tool-use loop over a real filesystem.
AnthropicCursorGitHub
200
Tabular ML & data

Building a Customer-Segmentation Engine

Beginner· ~1h 28m
🛠 Build a customer-segmentation engine that turns a table of customers into named, actionable segments
The engine behind the "segments" / "audiences" features in marketing & CRM analytics — Spotify, Amazon, Airbnb, HubSpot, Klaviyo — and behind customer analytics everywhere.
SpotifyAmazonAirbnb
📄 1 paper50

An internship, not a course.

The difference shows up on your CV, not on a certificate.

A courseA Learn {} Fast internship
What you doWatch, then quizBuild, then defend it
Organised byTopics — “Module 7”Outcomes — “Build X”
Source materialA summary of a summaryThe original papers
When you're stuckA forum, two days laterAskThili, the same second
You end withA certificateA system you built — and can explain

Meet the mentor who never leaves.

AskThili answers from the section you're on and the code in front of you — never a hallucinated tangent.

AskThili· reading §4 — Training the discriminator
Grounded in your section · not the whole internetSee how it works →

A score you can put in front of a recruiter.

Every internship ends in a real assessment — of the thing you built, not the hours you sat through.

2.4/3
Capability score
Two halves: what you know, and what you shipped.

A completion test written against that internship's own material, plus a rubric review of the repo you actually pushed.

Three levels of thinking, not one.

Questions span recall, application and creation — so the number reflects whether you can use the idea, not whether you recognise it.

Deterministic, and the same every time.

The score is arithmetic we own and can replay from the stored evidence — not a language model's opinion on the day. Same work, same number, explainable line by line.

It compounds.

Scores roll up across every internship you finish, weighted by difficulty, into one capability number that travels with you.

Watch your understanding take shape.

Every question you ask lights up a concept. Your Mind map shows what you've mastered — and what's next.

transformersattentionCNNsRLRAGGANs

Pay for what you learn.

No subscription you forget about. Energy powers your internships — top up only when you want more.

175 free to start — enough for your first beginner or intermediate internship, end to end
From 50 for a beginner internship · 200 for an advanced one
No lock-in, no subscription. The completion test is always free.

Pick one. Build something real this week.

Look around first — you don't need an account to browse.