Reading about applied AI doesn't get you hired or shipped — ten complete builds, each with real evals, real costs, real deployment.
You've read the papers and finished the tutorials, but your portfolio is still a folder of notebooks that only run on your laptop. This course is ten complete builds, each one a system you could put in front of real users. You'll ship a hybrid retrieval-augmented generation (RAG) search over your own documents, extend it to answer from diagrams and screenshots, and let non-technical teammates query a database in plain English without risking a dropped table. Then the operations layer around all of it: a pipeline that generates its own eval cases, a semantic cache that cuts your model bill, and an extraction service that scores its own confidence and routes doubtful fields to a human. Then agents — an orchestrator directing specialist workers, a browser agent that clicks through sites with no API, and a voice agent that answers a phone call in under 500 milliseconds. You finish by fine-tuning a small model with LoRA (low-rank adaptation) and proving it beats a managed API on cost per correct output. Every module ships with its own eval bar, deliverables, and rubric.
Built by Lakshya Kumar
Paste this into any AI chat. Fill in the bracketed parts with your context — you'll get back a straight answer on whether this belongs on your plate.
We grant free access case-by-case — students, career-switchers, builders on a tight budget. Sign in to send us a note.
Sign in to applyFinished the tasks? Take the prompt to your AI and get tested on it. We copy the prompt and open the app — just paste it in.
Non-technical teammates keep queueing behind you for numbers; give them plain-English querying that can't drop a table or leak a row.
Hand-writing eval cases stalls at fifty; build a pipeline that mines your logs and docs for hundreds you'd never have thought of.
Most production traffic asks the same handful of questions in different words — stop paying full price to answer each one again.
Manual data entry from PDFs is slow and quietly wrong; build extraction that scores its own confidence and flags what needs a human.
One agent holding twenty tools loses the thread — split the work across specialists and keep every context window clean.
Plenty of systems you need to drive have no API, only a login page; build an agent that navigates them safely and leaves an audit trail.
Voice forgives nothing: a 900ms pause feels like a dropped call, so build for sub-500ms and prove it against 30 scenarios.
When one narrow task dominates your API bill, a small model you tuned yourself can beat the frontier on cost per correct output.
Complete all modules, then submit the required number of capstone projects. Each must earn a passing rating from an admin reviewer.
Pick 3 of the 10 projects that match your current work or career interest. Ship each to production-quality (real users OR realistic eval; documented deliverables per the module rubric). Submit: live links/repos for each, eval results, the 1-page writeup per project.
Ship Modules 1 (Hybrid RAG), 2 (Multi-modal RAG), and 3 (Text-to-SQL) together as a unified knowledge-search platform. All three over the same corpus where applicable. Submit unified demo + eval.
I'm taking "Applied AI Projects" — the capstone course of the Agentic and Applied AI track. 10 modules, each a full production project. Topics: hybrid RAG, multi-modal RAG, text-to-SQL with guardrails, eval generation, semantic caching, document extraction, agent orchestration, browser agents, voice agents, LoRA fine-tuning. My context: 1. My current product / project is: [describe] 2. My current AI experience: [shipped X / built Y / just starting] 3. My team's main pain right now: [hallucinations / cost / latency / safety / scope] 4. My biggest skill gap: [which of the 10 areas] Given that, answer: - Which 3 projects should I prioritize and why? - What's the realistic 4-week plan if I have 10 hours/week? - Name 1 project I should skip and why. - If I had to pick ONE project that would have the biggest impact on my work, which?
Ship Modules 4 (Eval Generation), 5 (Semantic Cache), and a 'monitoring + cost optimization' bundle as one cohesive ops layer on an existing LLM service. Submit code, dashboards, cost-savings report.
Ship Modules 7 (Orchestration), 8 (Browser), 9 (Voice) as a single multi-modal agent platform — text-driven orchestration of browser actions + voice handoff. Submit demo, traces, eval.
Ship Module 6 (Document Extraction) + Module 10 (LoRA Fine-Tuning) as a paired system: extract documents, label the extractions, fine-tune a small model on the labels, serve at lower cost. Submit pipeline + cost-saving analysis.
Module 8's foundation.