You built the AI project. It never shipped.

Someone asked what it costs at scale, or whether the retrieval was actually correct. There was no answer, so it stayed in the repo.

Try Week 1 at $75 See all 17 weeks

Enrolment closes October 16. Week 1's live build day is October 23. The $75 comes off the full fee.

Cohort 1 is running now

See the work students are shipping, week by week. Repos available on request.

eval run — support-assistant v4 golden set: 120 cases
pass retrieval@5 returns the source doc118/120
pass refuses when no source matches114/120
pass p95 latency under 2000ms1840ms
pass no PII in logged completions120/120
?? the answer is actually correctno test
Most candidates get to line four. The job is line five.
17Weeks
33Real builds
170+Hours
40Seats

// Recorded weekdays · Live weekend builds · Built for developers who already ship code

THE GAP

Why it died in pilot.

Almost never the model. Almost always the three things nobody teaches.

5%

of custom enterprise GenAI tools make it to production.

MIT Project NANDA, The GenAI Divide: State of AI in Business 2025. RAND (2024) cites estimates that AI projects fail at twice the rate of non-AI IT projects.
01

Evaluation

No golden set, no adversarial inputs, no harness running on every change. So nobody could say whether the citations were real.

02

Cost

No token accounting, no caching, no small-model routing. The number at ten times the traffic was a guess, and the guess was frightening.

03

Reliability

No tracing, no fault injection, no answer for what happens when the vendor rate-limits you at 2am.

WHAT YOU SHIP FIRST

Week 1, both builds. Read the code yourself.

One you write alongside the instructor, line by line. One you ship on a new input and an engineer scores it.

ReleaseBot

guided · live day FastAPI · Pydantic · SSE

RAW RELEASE NOTES IN

- api is way faster now (~40%)
- killed the old v1 api
- fixed that annoying login bug
- added dark mode i guess
- db migration, run it before deploy

STRUCTURED CHANGELOG OUT, STREAMING

PERF  Request latency reduced ~40%
      p95 340ms → 204ms
BREAK v1 API removed
      migration required before deploy
FIX   Authentication session bug
FEAT  Dark mode
OPS   Database migration
Notice what it did. "api is way faster now" became a PERF entry with p95 numbers attached. "killed the old v1 api" was classified BREAK and linked to the migration. That is semantic classification, not string formatting — Week 7 is where you prove it holds on inputs it has never seen.

MinuteMaker

you ship it · graded scorecard returned

MEETING TRANSCRIPT IN

SARAH: ok so the vendor thing. we can't
sign until legal clears it.
RAJ: i can chase them, probably end of
next week?
SARAH: fine. and someone needs to tell
the client we're slipping.
RAJ: that's you, you own that relationship.
SARAH: ...yeah ok.

STRUCTURED MINUTES OUT

DECISION
Vendor contract blocked pending legal review

ACTIONS
[Raj]   Chase legal on vendor clearance
        due 2026-09-18
[Sarah] Notify client of timeline slip
        owner inferred from context
The hard part is not extraction. Nobody said "Sarah owns this action" — she said "…yeah ok." Inferring ownership from a conversational concession, and knowing when you are wrong to infer it, is the difference between a demo and something a team would use.
HOW IT COMES BACK

Then an engineer reads it and scores it.

Not a rubric checkbox and never self-assessed. This is a real Week 1 scorecard, returned before Week 2 opens.

WEEK 1 · GRADED

MinuteMaker

78/100
Correctness
22/25
Completeness
21/25
Design
16/25
Clarity
19/25
Reviewed by an engineer, not a bot. Returned before Week 2 opens.
PROOF

Cohort 1 shipped these. They are running now.

Two services built by students in Cohort 1, deployed and responding. Repos available on request.

MinuteMaker

Built by Pranav in Cohort 1. MinuteMaker turns meeting transcripts into structured meeting minutes and action items, then delivers the summary directly to Gmail.

FeedbackSorter

Built by Ankush Saha in Cohort 1. FeedbackSorter classifies incoming feedback by intent and urgency, then benchmarks model performance to identify the best fit for the task.

17 WEEKS IN FULL

Sixteen teaching weeks. One capstone week.

Open any week to see what you learn, the stack you use, and the two projects that leave your machine at the end of it.

  • Onboarding + fast-track foundations (no prior AI assumed)
  • How modern AI products are built, layer by layer
  • LLM internals and how RLHF shapes behaviour
  • The model / retrieval / tool / memory stack
  • Streaming responses and structured output
  • Latency, cost and quality tradeoffs
PythonFastAPIPydanticOpenAI / AnthropicSSE
PROJECT 1 · BUILT LIVE
ReleaseBot — A service that streams messy notes into a clean, structured changelog.
PROJECT 2 · GRADED
MinuteMaker — A service that turns a meeting transcript into structured minutes and action items.
  • How LLMs actually work under the hood
  • Choosing the right model and provider
  • An evaluation primer: measuring quality
  • Provider tradeoffs across the market
  • Cost and latency-aware model selection
OpenAIAnthropicOpen-source modelsPython
PROJECT 1 · BUILT LIVE
IntentIQ — An intent classifier that picks the right model for the job.
PROJECT 2 · GRADED
FeedbackSorter — A tool that sorts inbound feedback by intent and urgency.
  • System prompt design and few-shot patterns
  • Strict output schemas with Pydantic
  • Function and tool calling, end to end
  • Streaming (SSE) patterns
  • Classification and routing logic
Pydantic v2Tool callingFastAPISSE
PROJECT 1 · BUILT LIVE
TicketStream — A bot that classifies and routes incoming requests automatically.
PROJECT 2 · GRADED
ReviewRouter — A bot that tags product reviews by sentiment and topic, then routes them.
  • Embeddings and cosine similarity
  • Chunking strategies and their tradeoffs
  • Vector databases and indexing
  • Multimodal ingestion — PDFs, images, tables
  • Semantic retrieval by meaning
EmbeddingsVector DBPython
PROJECT 1 · BUILT LIVE
KnowledgeVault — A search engine over your documents, by meaning not keywords.
PROJECT 2 · GRADED
PaperFinder — A semantic search tool over a set of research papers.
  • The full RAG pipeline, end to end
  • Grounding answers in retrieved sources
  • Generating trustworthy inline citations
  • Measuring whether an answer is grounded
  • Fallback behaviour when no source fits
RAG pipelineVector DBOpenAI / Anthropic
PROJECT 1 · BUILT LIVE
CitationRAG — An assistant that answers from documents and cites every claim.
PROJECT 2 · GRADED
DocuRAG — A cited RAG assistant over a document set of your choice.
  • Query transformation — HyDE, step-back
  • Reranking and context compression
  • Context engineering principles
  • Conversation state management
  • Improving retrieval measurably
HyDERerankingVector DBPython
PROJECT 1 · BUILT LIVE
RAGOptimizer — An upgraded assistant that improves retrieval and proves the lift.
PROJECT 2 · GRADED
RerankLab — A lab that applies reranking and compression to a new corpus.
  • Building golden test datasets
  • Pass/fail thresholds and metrics
  • Adversarial testing and red-teaming
  • Experiment design for AI systems
  • Turning results into a real report
Golden datasetsEval harnessPython / Pandas
PROJECT 1 · BUILT LIVE
BreakRAG — A harness that stress-tests a RAG system to find where it breaks.
PROJECT 2 · GRADED
RAGBench — A benchmark that scores a RAG system before and after on a golden set.
  • When to fine-tune versus prompt
  • LoRA and parameter-efficient fine-tuning
  • Running open-source LLMs
  • Small language models (SLM)
  • Benchmarking a tuned model
LoRA / PEFTOpen-source LLMsPython
PROJECT 1 · BUILT LIVE
SpecialistTuner — A fine-tuned small model specialised for one task.
PROJECT 2 · GRADED
DomainTuner — A small model fine-tuned on a domain dataset, benchmarked against the base.
  • Workflows versus agents, and when to use each
  • The agent loop from first principles
  • ReAct, reflection and planning patterns
  • Agent state management
  • When NOT to use an agent
Agent loopTool usePython
PROJECT 1 · BUILT LIVE
OpsAssist — An agent that completes a multi-step task on its own using tools.
PROJECT 2 · GRADED
ResearchAgent — An agent that researches a document set and writes a cited brief.
  • Multi-agent orchestration patterns
  • Router and specialist agent designs
  • Long-term and shared memory
  • Inter-agent communication
  • Coordinating a multi-step workflow
Multi-agentMemory storePython
PROJECT 1 · BUILT LIVE
TriageFlow — A system that coordinates multiple agents to resolve work.
PROJECT 2 · GRADED
DeskOrchestrator — A router-plus-specialist agent pipeline over a workflow.
  • The Model Context Protocol (MCP)
  • Agent-to-agent (A2A) communication
  • AGENTS.md and tool discovery
  • Exposing tools to agents safely
  • Building interoperable agent systems
MCPA2APython
PROJECT 1 · BUILT LIVE
AgentMesh — A mesh that connects agents and tools across an interoperable layer.
PROJECT 2 · GRADED
ToolBridge — An MCP server that exposes a set of tools to an agent.
  • Prompt injection attacks and defence
  • PII detection and redaction
  • A practical responsible-AI checklist
  • Human-in-the-loop checkpoints
  • Bias, fairness and safety basics
PII detectionGuardrailsPython
PROJECT 1 · BUILT LIVE
GuardianAI — An assistant hardened with guardrails and PII protection.
PROJECT 2 · GRADED
SafeAssist — A guarded assistant for a sensitive domain with a human in the loop.
  • Building developer-tooling agents
  • Streaming UI patterns
  • UX feedback and progressive disclosure
  • Agents that read and act on code
  • Designing agents for trust
Streaming UIAgentsPython
PROJECT 1 · BUILT LIVE
WorkbenchAI — A developer-tooling agent with a live, streaming interface.
PROJECT 2 · GRADED
PRCopilot — An agent that reviews a pull request or summarises a diff, live.
  • Containerising AI services with Docker
  • Deploying to a public URL
  • Observability: tracing and metrics
  • Versioning and rollbacks
  • Async request handling
DockerTracing / metricsCloud deploy
PROJECT 1 · BUILT LIVE
DeployCore — An AI service deployed the production way, live to a public URL.
PROJECT 2 · GRADED
ObserveOps — Tracing, metrics and versioning added to a deployed service.
  • Retries, fallbacks and timeouts
  • Circuit breakers for AI calls
  • Fault injection and chaos testing
  • Graceful degradation
  • Producing a resilience report
Reliability patternsFault injectionPython
PROJECT 1 · BUILT LIVE
ReliabilityKit — A toolkit that makes an AI service resilient under failure.
PROJECT 2 · GRADED
ChaosProbe — A probe that fault-injects a service and reports its resilience.
  • AI FinOps: measuring and cutting cost
  • Prompt caching strategies
  • Small-model routing
  • Self-healing patterns
  • Producing a cost-savings report
Prompt cachingSLM routingPython
PROJECT 1 · BUILT LIVE
CostGuard — A layer that cuts an app's AI spend without losing quality.
PROJECT 2 · GRADED
CachePilot — Caching plus small-model routing added to an app, with a savings report.

The whole course comes together in one real production system — not a demo build. Choose one of six ready-made capstone projects, each shipping with a cleaned, licensed dataset, or bring your own data and we help you source and scope it.

WHAT YOU DELIVER
A deployed capstone system, a live demo and architecture walkthrough, and 33 builds on your GitHub.
HOW IT IS GRADED
A panel reviews your build and your presentation, scored on the same four criteria, with detailed written feedback.
THE CAPSTONE

Week 17: a system you present and defend.

Every track ships with a cleaned, licensed dataset. Or bring your own idea and we help you scope it.

CaseCompass

Ask 40,000 real court opinions anything: grounded, cited answers over public US case law.

LEGAL

MediGuard

A patient-info assistant with medical-grade guardrails, grounded in official health sources.

HEALTHCARE

FilingScout

An analyst agent that reads SEC filings and writes cited earnings briefs on any company.

FINANCE

StackSage

A developer support copilot trained on real Stack Exchange Q&A and open-source docs.

DEV TOOLS

TravelGenie

An agent that plans complete day-by-day itineraries, cited from real travel guides.

CONSUMER

PulseBrief

A scheduled agent that turns the week's world events into one clean, cited briefing.

MEDIA & OPS
WHO TEACHES IT

Meet the course designers and instructors.

Two practitioners, not professors. Every credential is public and linked — check them before you read anything else.

Randeep S. Bhatia

Randeep S. Bhatia

CTO AT SPLASH
EX TWITCH, AUDIBLE & EA · AAAI PUBLISHED

Advises on architecture and scale, drawing on a career shipping systems used by millions at Twitch, Audible and EA.

LinkedIn ↗
Vinay Bamil

Vinay Bamil

COURSE DESIGNER · PHD IN AI
EX GEN AI COACH, GOOGLE

Designed the course and teaches most of it, bringing research-grade understanding of LLM internals to a curriculum built entirely around shipped work.

LinkedIn ↗
INDUSTRY VALIDATED

Two leaders put their names to it.

Senior technology leaders who read the curriculum and went on the record.

Most AI courses produce people who can talk about AI. CoreSmart produces people who can actually ship. Every week delivers a usable artifact rather than just a certificate.

Ashish MagoCo-founder & CTO, Compliance Kart · Ex CTO, LMCE · Ex Director Technology, Publicis Sapient LinkedIn ↗

Teams already have AI tools, but delivery velocity remains unchanged. CoreSmart's model addresses the real gap by helping developers govern AI systems at scale.

Jignesh ModiCXO, Global Hospitals & Healthcare Technology · Digital Transformation & AI Leader LinkedIn ↗
PRICING

Apply for Cohort 2.

It costs $75 because it is a real week of the course, not a sales page.

WEEK 1
$75
Comes off the full fee if you continue. Not refundable on its own.
  • Prep week, unlocked within 24 hours
  • Week 1 in full — ReleaseBot and MinuteMaker
  • Your graded scorecard, returned by an engineer
  • Full 17-week syllabus, downloadable
Try Week 1 at $75
FULL PROGRAMME · COHORT 2
$1,999$1,499
Early bird for this cohort. List price applies from Cohort 3.
  • 17 weeks · 33 deployed builds · capstone
  • Every graded project reviewed by an engineer
  • Live weekend build sessions, recorded weekdays
  • Or three monthly instalments of $500
  • Your $75 deducted from this figure
Enrol now — $1,499

// 40 seats · Cohort 2 · Enrolment closes 16 October

Post payment, LMS login via email in 24 hours. Enrolment closes October 16; Week 1's live build day is October 23, and you need to be enrolled to be in that session.

QUESTIONS

Before you ask us.

Why pay for this when Claude can write the pipeline?

It can, and you should let it. What it cannot do is tell you whether the retrieval is correct, what the system costs at ten times the traffic, where it fails under load, or defend the architecture to a panel deciding whether to fund it. Only 29% of developers say they trust AI output to be accurate. Closing that gap is judgment.

How much time does it take?

Around 10 hours a week, roughly 170 across the 17 weeks. Weekday content is recorded. Build sessions are live on weekends and recorded, so missing one does not put you behind.

What does the $75 get me, and is it refundable?

Prep week and Week 1 in full — both builds and your graded project returned with a scorecard and written feedback from an engineer, plus the complete syllabus to download. It is not refundable on its own; it is credited in full against the fee if you enrol. It costs $75 rather than nothing because people who pay finish.

Do I need an AI background?

No. Week 1 opens with fast-track foundations including a Python and API refresher. You should be comfortable writing code and working with APIs. You do not need ML going in.

Why should I trust a new programme?

Do not take our word for it — check the people. Randeep's Twitch and Audible record, and Vinay's Google and PhD credentials: both public, both verifiable in a few minutes from the links above. Ashish Mago and Jignesh Modi have publicly attached their names to the curriculum. And the two Cohort 1 services on this page are running right now, and the repos are available on request.

What is the refund policy on the full fee?

Full refund within 14 days of enrolment, no questions asked. After 14 days the fee is non-refundable, since by then you have had the live sessions, the graded reviews and the material. The $75 is separate: it is not refundable on its own, and it is credited in full against the fee if you enrol.

Who reviews my work?

An engineer, every week, against correctness, completeness, design and clarity. Written feedback, never a bot and never self-assessed.

Try Week 1 at $75Cohort registration closes 16th Oct
Try Week 1 →