SousChef
FastAPI + Streamlit recipe system with AI extraction from text/images/URLs, JWT auth, Redis-backed background work, Docker Compose and Playwright E2E.
EASS
Engineering of Advanced Software Solutions is a two-part HIT course about end-to-end ownership. Part A: one student owns a software system from problem to deployment. Part B: a team expands that ownership to users, operations, measurement, and distribution.
Own the whole system.
Each student takes an idea to a running, testable product and practices architecture, interfaces, reliability, deployment, and end-to-end ownership.
Problem→Architecture→API + data→UI→Tests→Deploy→E2E demo
Core: Python & uv · Git/Linux · HTTP/APIs · FastAPI · persistence/migrations · frontend · Docker/Compose · testing/CI · auth/security · async work · caching/queues · cloud/AI · observability.
Own the operating loop.
Teams choose a real problem, find users, ship, operate the product, measure what happens, and iterate from evidence. Product, engineering, ML, reliability, and GTM become one operating loop.
Problem→Users→MVP→Ship→Measure→Talk to users→Iterate
Core: problem selection · customer discovery · product scope · team ownership · production operations · analytics/experiments · AI/ML where useful · reliability/security · pricing/positioning · distribution/GTM · weekly shipping · metrics · demo day.
Selected Part A systems
A verified sample across cohorts showing progression from full-stack foundations to AI-native systems.
Prompt Vault
FastAPI prompt library with SQLModel/Alembic, Redis, a separate analyzer service, Claude via Pydantic AI, Docker Compose and async refresh.
Israel Flights Tracker
Scheduled flight-data ingestion with FastAPI, Streamlit, PostgreSQL, analytics endpoints and Docker Compose.
Notes App
React frontend, FastAPI backend, MongoDB and Docker Compose—the early-course end-to-end pattern in a compact system.
Selected Part B ventures
These examples show company-level ownership in practice: users, product, operations, measurement, iteration, distribution, and the code that supports them.
Fridgi
AI home-food management for inventory, recipe suggestions and shopping. The team also did customer segmentation, market research, revenue modeling, lead generation and outreach.
System → users → market → leads.
Prizma News
A live publishing product with content, readers, distribution, and an operating public surface.
Build → publish → operate → learn.
Waste-contamination detection
An earlier cohort applied computer vision to contamination in waste streams and connected model output to an operational decision.
From software systems to AI systems
AI is not a third course phase. It deepens the same ownership model: inspect the mechanisms, build a model you can understand, learn the full training system, then engineer the agent around it.
Mechanisms→Inspectable model→LM stack→Agent system
Implement the ideas, not just the papers
Sutskever 30 turns a foundational deep-learning reading list into 30 educational, NumPy-first Jupyter implementations—from RNNs, LSTMs, ResNets and attention through scaling laws, RAG and long-context behavior. Use it to connect paper-level ideas to code you can inspect, run, and modify.
Make the black box small enough to open
GuppyLM is an open-source, 8.7M-parameter Transformer trained from scratch on synthetic conversations. Its project traces the path from generated data and tokenizer through training, inference, and a browser demo. It is a teaching example, not a production model.
Engineer the full language-model stack
Stanford CS336 follows the complete path from tokenization and Transformer implementation through GPU kernels, parallelism, scaling, data, evaluation, and post-training, including reinforcement learning for reasoning.
Engineer the system around the model
Stanford CS329Z moves from retrieval, tool use, and memory to multi-agent systems, evaluation, safety, and coding agents. It makes the model one component inside a larger system whose behavior must be measured and reproduced.
Classroom connection: the components change, but the standard does not: define behavior, make the system runnable, evaluate it over meaningful tasks, inspect failure modes, and preserve reproducibility.
Course infrastructure
The course is supported by reusable practice and assessment systems. Some are public; some are internal course tools.
EASS Engineering Mastery
455 questions across 34 modules covering durable systems mental models and practical engineering.
EASS Auto-Grader
Parallel multi-repository evaluation with structured reports, code-similarity analysis, and originality checks for reproducible assessment.
EASS Hub
Internal course workspace for organizing engineering resources and course material.
Teaching thesis
Part A teaches individual technical ownership. Part B expands that ownership to a team, users, operations, metrics, and distribution. AI changes the components—not the standard of evidence.