वास्तविक सिस्टम जिन्हें आप बनाते और रखते हैं — प्रत्येक एक स्पेक, एक चलाने योग्य प्रमाण, और सटीक पंक्ति के साथ जिसे आप अपने रिज़्यूमे पर रख सकते हैं। एक बिल्ड चुनें, उसे शिप करें, सबूत के साथ आगे बढ़ें।
Every lab opens with a spec a real team would recognize — not a puzzle with one clever answer.
A starter scaffold and the Sage loop: map the system, decide under tradeoffs, ship the thing.
Each build ends in runnable acceptance checks — the proof a skeptic can re-run, the reason it belongs on your résumé.
Retrieve wide with embeddings, then re-rank with a real cross-encoder and measure the lift, not just eyeball it.
Cache by meaning, not exact string match — then prove it actually saves money without serving stale garbage.
Split the answer into claims, check each one against the retrieved context, and catch what the model made up.
Route the request to the right tool, validate every argument against its schema, and never let a malformed call through.
Accept third-party webhooks without trusting the network.
Paginate a list that keeps changing underneath you.
Speed up reads without ever serving stale data.
Stop bad data before it reaches the warehouse.
Windowed rendering that stays smooth at 10,000+ rows.
A dialog that keyboard and screen-reader users can actually use.
Contracts that fail the build the moment an API breaks a consumer.
A load test that fails the build when p99 latency blows the budget.
Deploy a green environment, gate traffic on real health checks, and roll back in one command.
Define roles and permissions once, enforce them at every route, and prove a lower-privilege user gets denied.
Escape and allowlist untrusted HTML at the boundary, then prove a real stored-XSS payload renders inert.
Convert a stack-overflowing recursive function into an iterative one with an explicit stack, and prove it survives inputs the original crashed on.
A small RAG service: you index a folder of documents into a vector store, retrieve the top-k passages for a question, and have the model answer strictly from that context — citing the source and abstaining when the answer isn’t present. You finish with a service that turns "the model guessed" into "the model looked it up."
A test suite for an AI feature: a golden set of real inputs with known-good answers, a scorer (rubric or judge) that grades each output, and a single aggregate score. You change a prompt and watch the number move — the way a real team ships model changes without breaking things.
A charge endpoint that stays correct when the network retries. You add an idempotency key, a store of handled keys, and a guard that makes a repeated request a no-op returning the original result — so three retries still charge the card once.
An agent that can act, but never fires the loaded gun without you signing off first.
Untrusted text is the new SQL injection — build the filter that catches it before it reaches your model.
Make the model hand you valid JSON every single time, not just when it feels like it.
Embeddings are cheap to generate and easy to get wrong — prove yours actually retrieves the right thing.
Per-key token buckets that cap burst traffic without punishing steady, well-behaved callers.
At-least-once processing with exponential backoff and a dead-letter queue for the jobs that never recover.
The expand-backfill-contract pattern that lets old and new code read the same table during a live migration.
Extract, transform, load — designed so a re-run, a retry, or a crash mid-load never duplicates a row.
No headless library. Just you, ARIA, and a screen reader that has to actually understand your widget.
Instant-feeling UI is easy. Instant-feeling UI that doesn't lie to the user when the server says no is the actual skill.
Most E2E suites are flaky theater. Prove yours isn't by planting a real bug and watching it fail.
Pixel-diffing is trivial. A harness that doesn't cry wolf on font antialiasing is the hard part.
Define your infra as code, deploy it for real, then prove you can roll back to the exact prior state.
Move a hardcoded secret into a real secret store, then rotate it live with zero downtime.
Password hashing, sessions, CSRF protection, and rate limiting — then prove the common attacks fail.
A program is silently wrong. Instrument it, isolate the exact failing line, explain why, then lock in the fix with a regression test.