AI Engineering — 4-Month Plan

🎯 The product you're building

ReviewLoop — get more 5-star Google reviews on autopilot. A WhatsApp AI that has a real conversation with a business's customers, sends happy ones to Google and unhappy ones privately to the owner, in their own language — and turns the feedback into insight.

Patient visits → WhatsApp reaches out (right language) → AI has a short adaptive conversation → HAPPY: one-tap Google review link | UNHAPPY: private recovery + instant owner alert → every chat analyzed → owner sees rating climb + WHY, in a dashboard

Beachhead
UAE health-&-beauty clinics (dental, derma, med-spa). Reviews are existential; owners pay $99–149/mo; you can sell hand-to-hand in Dubai. Product is global-capable — only the first ~30 customers are local.

The wedge
Conversational AI interviewer (not a form) · WhatsApp-native (incumbents are email/US-SMS) · multilingual · SMB-priced insights · compliant-by-design routing (no illegal review gating).

Full detail in product-brief.md. Validation & sales playbook in go-to-market.md. This page is the learning plan (source: curriculum.md).

🎥 Video courses & follow-alongs

The learning loop: watch a course to learn it → follow a build tutorial → use the docs to reference it. Every week below also lists its docs; the videos live here and in curriculum.md. All links verified live (2026).

⭐ Start here — spine your first two months

course · freeBuilding with the Claude API — Anthropic Academy · ~8h — API, prompting, tool use, RAG, agents. The single best course for your core skill. (Coursera + certificate)

🏗️ Watch one end-to-end build early

tutorial · freeShip a Production RAG App w/ FastAPI — freeCodeCamp · closest to your architecture
video · freeAgentic AI in 10 Hours — Krish Naik
course · freeLearn GenAI in 30 Hours — freeCodeCamp
video · freeBuild & Deploy a SaaS AI Agent Platform

📺 Channels to subscribe to

channelDave Ebbelaar · production AI in plain Python
channelCole Medin · production agent/RAG builds, all code on GitHub
channelSam Witteveen · agents, evals, production
channelAI Jason · multi-agent systems, architecture

By topic

Python & async

FastAPI

RAG

Agents

WhatsApp integration

Evals & guardrails

Observability

Docker & deploy

On-device / MLX (stretch)

🐍 Phase 0 — Learn Python first ~2 weeks, before Week 1

You're a senior programmer, not a beginner — this is "Python for people who already code," not intro-to-coding. Move fast; fluency comes as you build.

⭐ Recommended — learn fast + earn a free credential

1 · freeLearn Python in Y Minutes · ~1h — map Swift→Python syntax
2 · free certCS50P — Harvard Intro to Programming with Python · fast, idiomatic, free Harvard certificate (doubles as your first cert)
3 · freeExercism Python track + Codewars — build fluency by doing, with mentoring

Pure speed (no cert)

freeOfficial Python Tutorial
freeGoogle's Python Class · for devs who know another language

Video / free browser cert

freeMosh — Python Full Course (2025, ~2h) at 1.5×
freeCorey Schafer — Python idioms playlist
free certfreeCodeCamp — Scientific Computing with Python

⏱️ Adds ~2 weeks. Front-load it at higher intensity to hold the 4-month cap. Then Week 1 picks up with modern Python idioms + async + Pydantic.

🗓️ Month 1 — Foundations & your first LLM calls Weeks 1–4

Python + async + FastAPI + first Claude calls. Lean on your Swift instincts: Pydantic ≈ Codable/structs, async/await ≈ Swift Concurrency.

Week 1 — Modern Python for a Swift dev

Type hints · async/await + asyncio · uv environments · Pydantic v2.

docsuv — Official Docs
docsPydantic v2 — Official Docs
docsasync/await — clearest plain-English explainer
articleAsync IO in Python — Real Python
articlePython Type Checking — Real Python

Build: repo with uv; typed Pydantic models IncomingMessage / Conversation; a small async script.

Week 2 — FastAPI backend

Async endpoints · Pydantic I/O · webhooks · background tasks.

docsFastAPI Tutorial — User Guide
docsBackground Tasks — FastAPI
docsOpenAPI Webhooks — FastAPI
videoFastAPI for Beginners — Full Course (2025)
videoPython API Dev — freeCodeCamp (19h)

Build: FastAPI /webhook that accepts a mock WhatsApp payload (Pydantic-validated) and logs it.

Week 3 — The Claude API & prompting

Messages API · prompt engineering · streaming.

docsMessages API Reference — Claude Docs
courseAnthropic Interactive Prompt Engineering Tutorial
cookbookAnthropic Claude Cookbooks

Build: webhook calls Claude Sonnet 5 to generate a warm reply; iterate the review-ask prompt.

Week 4 — Tool use, structured outputs & prompt caching

Function calling · forced JSON · caching the system prompt.

Build: classify sentiment (happy|neutral|unhappy) via structured output with Haiku 4.5; cache the per-clinic system prompt.

🏁 Milestone 1: a FastAPI service that receives a (mock) WhatsApp message, generates a Claude reply, and classifies sentiment. Interview-showable.

🗓️ Month 2 — Product engine: RAG + agent + real integrations Weeks 5–8

Turn it into the actual review loop, wired to real WhatsApp + Google.

Week 5 — RAG & vector databases

Embeddings · chunking · retrieval · pgvector (prod) or Chroma (prototype).

freePinecone Learn: RAG (full-pipeline primer)
docspgvector — Postgres vector search
docsQdrant Documentation
docsChroma — Getting Started
courseRAG — DeepLearning.AI (Zain Hasan)

Build: RAG over a clinic's FAQ/services (Voyage embeddings + pgvector); answer "do you do teeth whitening? hours?".

Week 6 — Agents & tool-calling patterns

The tool_use loop · workflow vs agent (favor simple).

articleBuilding Effective Agents — Anthropic
cookbookAnthropic Cookbook: Agent Patterns
docsTool use with Claude — Overview
courseHugging Face AI Agents Course

Build: the conversational review agent — multi-turn, adaptive follow-ups, a tool to log outcome + route.

Week 7 — WhatsApp Cloud API integration

Cloud API setup · webhooks · send/receive · the free 24h window.

Build: connect a real WhatsApp number; wire the agent to real inbound/outbound messages.

Week 8 — Google Business Profile API

Read reviews · post owner replies · generate the review deep-link. (API can read + reply, not post reviews for users.)

docsWork with review data — Google Business Profile API

Build: pull reviews, draft AI replies (reviews.updateReply, one-tap approve), generate the "leave a review" link for happy customers.

🏁 Milestone 2 (the MVP core): real WhatsApp conversation → sentiment → happy get a Google review link / unhappy alert the owner → RAG answers questions. A demoable product.

🗓️ Month 3 — Production discipline: evals, guardrails, compliance Weeks 9–12

This month separates an AI engineer from a prompt-wrapper. Don't skip it — it's your biggest hiring signal and your product's trust layer.

Week 9 — Eval harness & golden datasets

Building evals · LLM-as-judge · golden sets · regression tests.

Build: ~30-conversation golden set; eval sentiment + routing accuracy; LLM-as-judge for conversation quality.

Week 10 — Guardrails, prompt injection & the compliance guardrail

Guardrails · jailbreak/injection defense · policy enforcement.

Build: enforce "always offer the public review to everyone" (no gating) as a tested compliance guardrail; add injection defenses.

Week 11 — Insight extraction + human-in-the-loop

Structured extraction · aspect-level theme/sentiment clustering.

Build: cluster themes across conversations into a dashboard ("Tuesday wait-time complaints ×6"); add a human-in-the-loop validation view.

Week 12 — Multi-tenancy & data isolation

Per-tenant schema · isolation · onboarding.

Build: per-clinic tenancy + data isolation; onboarding = connect Google Business Profile + WhatsApp number.

🏁 Milestone 3: an eval'd, guardrailed, multi-tenant product with an insights dashboard. The "real AI engineer" proof.

🗓️ Month 4 — Observability, multilingual, ship Weeks 13–16

Deploy, go multilingual, get pilot clinics live.

Week 13 — Observability & tracing

Tracing · cost/latency/quality dashboards.

Build: Langfuse/OTel tracing across tenants; a cost/latency/quality dashboard.

Week 14 — Multilingual (Arabic + English)

Auto-detect + respond in language · multilingual evals.

Build: Arabic/English auto-detection + responses; extend the eval harness to Arabic.

Week 15 — Deploy & harden

Docker · hosting · managed Postgres · secrets.

docsFastAPI in Containers — Docker
docsDeploy FastAPI — Railway
docsRun a FastAPI app — Fly.io
articleDeploy FastAPI + PostgreSQL on Render
docsSettings Management — pydantic-settings

Build: Dockerize; deploy to Railway/Render/Fly; managed Postgres; secrets; a security pass.

Week 16 — Onboard design partners + polish

Ship it to real clinics.

Build: onboard 3–5 pilot clinics; weekly WhatsApp/email digest; a simple landing page.

Optional stretch — On-device / MLX (your latent iOS moat)

Only after the core ships. Run sentiment/PII-redaction on-device to cut cost + a privacy story.

docsmlx-lm — Run LLMs with MLX
videoWWDC25: LLMs on Apple silicon with MLX
docsMLX Documentation
articleRun LLMs on macOS with llm-mlx — Simon Willison
docsCore ML Tools — Convert Models

🏁 Milestone 4: a deployed, multilingual, observable MVP live with pilot clinics + a validated pricing test. Your portfolio centerpiece and a real business.

📜 Certifications — honest & tiered

The truth first: a shipped RAG/agent project (this plan) beats any certificate for an AI/LLM eng role — the strongest candidates hold the fewest certs. Treat certs as a screen-passer/tiebreaker, not the main event. Full detail in certifications.md.

🥇 Tier 1 — worth-it exams (pick ONE, near the end)

$200 · best fitDatabricks GenAI Engineer Associate · RAG, embeddings, agents, evals
$150AWS ML Engineer – Associate · strong ATS keywords (fintech=AWS)
$165Azure AI Engineer AI-102 · only if targeting Azure
$125NVIDIA NCA-GENL · cheap brand badge, optional

🆓 Tier 2 — free; the PROJECT is the value

free certHugging Face AI Agents · cert needs a working agent (Week 6)
freeAnthropic — Building with Claude API · paid Coursera cert optional
free certHarvard CS50P · doubles as your Python on-ramp

Skip / don't stack: prompt-engineering "get hired" badges, Google GenAI Leader (business, not eng), Azure AI-900, Kaggle micro-certs — fine as free learning, not resume headliners.

Your move: free HF Agents cert + CS50P along the way; optionally one paid Databricks exam near the end. Don't over-invest in cloud-ML certs — you build on Claude + your own stack.

📣 Go-to-market (runs in parallel — Weeks 1–8)

Distribution, not code, is what kills solo founders. Do customer discovery while you build. Full playbook in go-to-market.md.

Validate first (cheap, before heavy build)

Interview 15–20 reachable Dubai clinics on how they get reviews + what they'd pay.
Wizard-of-Oz: run the WhatsApp loop by hand for 3–5 clinics; measure response lift + net-new reviews.
Landing page + real pricing; measure trial conversion.

Then sell (the sequence)

🚶 Walk-in pilots at clinic clusters (you're in Dubai — your unfair edge).
💬 Personalized WhatsApp/IG/LinkedIn DMs citing their current rating.
🏢 White-label to local SEO/marketing agencies (one deal = many clinics).
📝 Content/SEO to compound → 🎯 ads only once a funnel is proven.

🧰 Model & tooling choices (verified current, 2026)

Job	Choice	Why
Conversational agent + insights	`claude-sonnet-5`	Best speed/intelligence; intro $2/$10 per Mtok (thru Aug 31 2026)
High-volume sentiment/classification	`claude-haiku-4-5`	Fast + cheap ($1/$5)
Occasional hard synthesis	`claude-opus-4-8`	Most capable ($5/$25)
RAG embeddings	Voyage AI `voyage-4`	Anthropic has no first-party embeddings; recommends Voyage
Vector storage	pgvector (or Chroma)	Reuse the Postgres you already need; Chroma for quick local protos
Cost control	Prompt caching	Cache the per-clinic system prompt

Verified platform facts: WhatsApp inbound replies within the 24h window are free (marketing/auth templates are charged). The Google Business Profile API can read reviews + post owner replies but cannot post a review on a customer's behalf — so happy customers get a Google review link. No review gating (FTC/Google).