Capstone Projects

Projects are the strongest signal in a placement process. This catalog gives buildable ideas that map to the other tracks. Pick by your target role. Each project page in the app tracks status (planned, building, shipped) and links a document and a drawing. Aim to ship at least one starter, one intermediate, and one advanced project per target role.

How to read each entry: Problem, Stack, Milestones, Difficulty, Proves.

Problem: answer questions over a set of PDFs with citations.

Stack: Python, FastAPI, an embedding model, Qdrant, an LLM.

Milestones: ingest and chunk, embed and store, retrieve and rerank, answer with citations, evaluate with RAGAS.

Difficulty: intermediate.

Proves: retrieval, chunking, evaluation.

Problem: research a topic, browse, synthesize, and write a report.

Stack: LangGraph, a web search tool, an LLM, FastAPI.

Milestones: planner, search tool, writer, editor, human approval gate, tracing.

Difficulty: advanced.

Proves: agent orchestration, tool use, human in the loop.

Problem: review a pull request and suggest improvements.

Stack: LangGraph, the git platform API, an LLM.

Milestones: fetch diff, analyze, comment, summarize, guardrails.

Difficulty: advanced.

Proves: multi step agents, structured output.

Problem: pull typed fields from invoices and receipts.

Stack: Python, an LLM with structured output, pydantic.

Milestones: schema, extraction, validation, batch mode, accuracy report.

Difficulty: intermediate.

Proves: function calling, validation.

Problem: search your own notes by meaning.

Stack: sentence transformers, FAISS, a small web UI.

Milestones: index, search, hybrid keyword and vector, ranking.

Difficulty: starter.

Proves: embeddings, hybrid search.

Problem: answer support questions and escalate hard ones.

Stack: RAG, an LLM, a ticket system, FastAPI.

Milestones: knowledge base, retrieval, confidence threshold, escalation, analytics.

Difficulty: intermediate.

Proves: RAG, routing, guardrails.

Problem: ingest financial events and write clean tables.

Stack: PySpark Structured Streaming, Delta Lake, object storage.

Milestones: source, transform, watermarks, Delta write, dashboard.

Difficulty: advanced.

Proves: streaming, lakehouse.

Problem: bronze, silver, and gold layers over raw data.

Stack: object storage, Auto Loader, PySpark, a catalog.

Milestones: ingest, clean, model, quality checks, governance.

Difficulty: advanced.

Proves: lakehouse architecture, data quality.

Problem: schedule and monitor a daily ETL.

Stack: Airflow or Prefect, Python, a warehouse.

Milestones: extract, validate, transform, load, alerting.

Difficulty: intermediate.

Proves: orchestration, reliability.

Problem: turn raw tables into tested marts.

Stack: dbt, DuckDB or a warehouse.

Milestones: staging, marts, tests, docs, a dashboard.

Difficulty: intermediate.

Proves: dimensional modeling, testing.

Problem: train, register, serve, and monitor a model.

Stack: MLflow, FastAPI, Docker, Evidently.

Milestones: pipeline, registry, serving, drift monitoring, retrain trigger.

Difficulty: advanced.

Proves: the full MLOps lifecycle.

Problem: predict customer churn and serve scores.

Stack: scikit-learn or xgboost, FastAPI, a feature store.

Milestones: features, model, evaluation, serving, monitoring.

Difficulty: intermediate.

Proves: tabular ML, serving.

Problem: specialize a small model on domain data.

Stack: transformers, PEFT and QLoRA, a cloud GPU.

Milestones: dataset, train with LoRA, evaluate, serve, model card.

Difficulty: advanced.

Proves: fine tuning, evaluation.

Problem: regression test prompts and models.

Stack: promptfoo or deepeval, a dataset, CI.

Milestones: golden set, metrics, CI gate, cost report.

Difficulty: intermediate.

Proves: evaluation, cost control.

Problem: shorten and redirect with analytics.

Stack: FastAPI, PostgreSQL, Redis.

Milestones: hashing, redirect, analytics, caching, rate limit.

Difficulty: starter.

Proves: API design, caching.

Problem: a reusable distributed rate limiter.

Stack: Redis, FastAPI.

Milestones: token bucket, sliding window, distributed state, tests.

Difficulty: intermediate.

Proves: algorithms, distributed state.

Problem: run background jobs reliably.

Stack: Redis or a broker, workers, FastAPI.

Milestones: enqueue, workers, retries, dead letter, dashboard.

Difficulty: intermediate.

Proves: queues, reliability.

Problem: live messaging with presence.

Stack: WebSockets, Redis pub and sub, PostgreSQL.

Milestones: sockets, rooms, presence, history, fanout.

Difficulty: intermediate.

Proves: real time, pub and sub.

Problem: multi channel delivery with retries.

Stack: a queue, FastAPI, email and push providers.

Milestones: channels, priority, retry, deduplication, templates.

Difficulty: intermediate.

Proves: messaging, idempotency.

Problem: process orders with events and sagas.

Stack: Kafka or Redis Streams, FastAPI, PostgreSQL.

Milestones: events, CQRS, saga, outbox, idempotent consumers.

Difficulty: advanced.

Proves: event driven architecture.

Problem: a small replicated key value store.

Stack: Go or Python, a consensus library.

Milestones: storage engine, replication, consistent hashing, leader election.

Difficulty: advanced.

Proves: distributed systems fundamentals.

Problem: the very app described in the docs folder.

Stack: Next.js, Supabase, BlockNote, Monaco, Excalidraw, React Flow.

Milestones: see docs/04_build_phases.md.

Difficulty: advanced.

Proves: full stack, product thinking.

Problem: a fast notes app with offline support.

Stack: Next.js, a block editor, a database.

Milestones: editor, folders, search, offline, sync.

Difficulty: intermediate.

Proves: editors, local first.

Problem: drag charts onto a dashboard.

Stack: Next.js, a charting library, a database.

Milestones: data sources, chart config, layout, sharing.

Difficulty: intermediate.

Proves: data UI, state.

Problem: a shared drawing canvas.

Stack: Next.js, a canvas library, real time sync.

Milestones: canvas, shapes, real time, export.

Difficulty: advanced.

Proves: canvas, real time.

Problem: a full pipeline for a sample app.

Stack: GitHub Actions, Docker, Kubernetes, Argo CD.

Milestones: build, test, image, GitOps deploy, rollback.

Difficulty: advanced.

Proves: delivery automation.

Problem: full metrics, logs, and traces for a service.

Stack: Prometheus, Grafana, Loki, OpenTelemetry.

Milestones: instrument, dashboards, alerts, traces.

Difficulty: intermediate.

Proves: observability.

Problem: reusable Terraform modules for a stack.

Stack: Terraform, a cloud.

Milestones: modules, remote state, CI validation, docs.

Difficulty: intermediate.

Proves: infrastructure as code.

Problem: upload results and get a significance verdict.

Stack: Python, statsmodels, a simple UI.

Milestones: input, tests, power, report.

Difficulty: starter.

Proves: experiment analysis.

Problem: recommend items from interactions.

Stack: implicit or surprise, FastAPI.

Milestones: data, model, evaluation, serving.

Difficulty: intermediate.

Proves: recommendation, evaluation.

Problem: forecast a metric and serve predictions.

Stack: prophet or statsmodels, FastAPI.

Milestones: features, model, backtest, serve.

Difficulty: intermediate.

Proves: forecasting.

Problem: from raw data to a decision dashboard.

Stack: dbt, a warehouse, Metabase.

Milestones: model, test, dashboard, write up.

Difficulty: intermediate.

Proves: analytics engineering, storytelling.

Targeting AI engineering: build 1, 2, and 13.

Targeting backend: build 15, 20, and 21.

Targeting data engineering: build 7, 8, and 10.

Targeting full stack: build 22 and 25.

Targeting MLOps: build 11 and 14.

Ship fewer projects with depth rather than many shallow ones. For each shipped project, write a one page summary covering the problem, the approach, the trade offs, and the results, and record a short demo.