The 2026 Roadmap: Production AI/ML Systems

AI Agents, Advanced RAG, NLP, RecSys, MLOps + Weekly Community Sessions

Hey there 👋

Hope you had a great start to the year!

Celebrated some Christmas before falling sick for a few weeks :)

Small recap:
We started Jam with AI from literally zero subscribers a year back. Today, we’ve grown to nearly 38,000 builders!

THANK YOU SOOO MUCH ❤️
2025 was our first year building publicly. We learned what resonates with you, and we figured out where we can add the most value.

So here’s what’s changing in 2026. We’re doubling down: deeper systems, production-grade implementations, and projects that replicate how real companies actually build and evolve AI/ML infrastructure.

At the same time, we’re focusing our effort on building a real vibrant community through REAL conversations! More on this below.

Our focus this year centers on two themes:

1. System Thinking - The gap between knowing how to build an ML model or AI system and knowing the production decisions that matter. When do you need a full MLOps system vs when is a simple ML pipeline enough? When should you use Docker? How do you speed up CI/CD pipelines?, etc. This is what separates engineers who build demos from engineers who build systems that last.

2. Sovereign AI - The shift toward local, on-premise, and privacy-first AI infrastructure. Companies are moving from cloud dependency to owning their AI stack. We strongly believe everyone will own their own personalized AI. This isn’t hype. This is the future, and we’re building it with you.

   
   Big changes are coming. Here’s the full roadmap. 👇

---

---

What Sovereign AI Means (and Why It Matters)

The AI landscape is shifting. Companies are moving from “rent everything from the cloud” to “own your AI infrastructure.” Privacy regulations, data sovereignty laws, cost constraints, and security requirements are pushing organizations toward local, on-premise AI systems.

Sovereign AI means:

• Running models locally instead of API dependency

• Privacy-first architectures that keep data in-house

• On-premise or hybrid deployments, not cloud-only

• Owning your AI stack, not renting it

We believe everyone will eventually own their own personalized AI - running on their infrastructure, trained on their data, serving their specific needs. This isn’t hype. Companies are already making this shift, and its good to know how to build for it.

Throughout our projects (especially RAG, Agents, and NLP fine-tuning), we’ll show you how to build systems that can run locally, deploy on-premise, and scale without expensive API dependencies.

---

NEW: Building a Real Community : Friday Discussion Sessions on Substack

System Thinking Doesn’t Happen Through Courses Alone

You can read 100 blog posts on ML or AI Systems. But real understanding comes from discussions. From someone asking “Why did you choose Postgres over MongoDB here?” and getting an answer based on production constraints, not textbook theory.

That’s why we’re introducing Friday Discussion Sessions - live, open conversations where we tackle real questions, break down complex topics, and learn from each other’s production experiences.

Mainly targeted so both you and we learn from each other’s experience!

When & Where:

• Most Fridays at 3 PM CET

• 1 hour on Substack Live

What We’ll Cover:

• AI/ML Career Decisions - Growth paths, role transitions, what actually matters

• Research Paper Deep-Dives - Breaking down papers that matter for production

• System Architecture Reviews - Real production diagrams, design decisions, trade-offs

• AI Tooling Reality Checks - What’s worth it, what’s hype, what works at scale

• Resume & Interview Prep - Landing roles, showcasing production experience

• Open Q&A - Whatever questions you bring to the table

• Guest lectures - Get more industry experts to talk on a particular topic

‼️ This Friday‼️ (23/01/2026):
We’re kicking off with Data Science Roadmap: Where It All Started - exploring how data science roles evolved and where they’re heading in 2026.

---

Monthly Paid Subscriber Deep-Dives

Separate from Friday sessions, we’re continuing monthly calls exclusively for paid subscribers on very specific technical topics with detailed walkthroughs.

Upcoming Deep-Dives:

• RecSys System Walkthrough - This Sunday! High-level architecture breakdown. We will send details to you soon :)

• MLOps System Walkthrough - Coming soon!

These are focused technical sessions where we go deep into production systems, not general discussions.

Note: During months when we’re running active course sessions (like AI Agents cohort), we won’t have additional monthly calls to avoid overwhelming everyone.

---

The 2026 Roadmap: Main Projects

First, let’s be clear:

We’ll continue sharing career tips, production experience blogs, AI/ML tooling breakdowns, architecture deep-dives, and all the reality-check content you’ve come to expect from us. That’s not going anywhere.

Continuing with our “The Mother of AI” Project and adding NLP series to it!

But these are the BIG projects for 2026 - the structured, hands-on, production-grade systems we’re building with you this year:

1. AI Agents: Production-Grade Systems

Why This Course?

Most AI agent tutorials teach you to build a chatbot that works on your laptop. We’re teaching you to build production-grade agent systems - developed locally, but built with the same patterns and constraints we use to serve thousands (or millions) of real users.

Why Us?

Shantanu built and deployed AI agents at HeyJobs with his team, now serving thousands of users. Biggest win? Solving the cost problem that kills most agent projects.

Shirin works on AI agent development at a fintech, building customer chatbots for millions of users. Her team/company was the first fintech in Europe to deploy AI agents in production. She’s lived the entire lifecycle: deployment, optimization, scaling, and everything that breaks along the way.

We’ve both seen what works and what doesn’t when agents meet real users and real constraints.

What Makes This Different?

We go deeper than anyone else:

• Real-world challenges: latency spikes, cost explosions, hallucination handling, tool failures

• How to architect systems that don’t break when one component fails

• Cost optimization strategies

• The messy reality of production - monitoring, debugging, iterative improvement

• Production patterns that actually work vs what sounds good in blog posts

What You’ll Build:

A complete production-grade AI agent system (developed locally) using LangGraph, covering:

• Planning & Reasoning - Multi-step task decomposition and execution

• Tool Use - RAG integration, API calls, database queries, and custom tools

• Memory Systems - Conversation history, user context, and state management

• Multi-Agent Systems - Coordinating specialized agents for complex workflows

• Production Patterns - Monitoring, error handling, cost tracking, and optimization

Format:

Self-paced learning:

• Written blogs covering core concepts and implementation details

• Deep-dive videos walking through the code and production patterns

Live support:

• Weekly Q&A sessions where we unblock your specific challenges and review your implementations

Access:

• Paid Subscribers: 50% discount

• Free Subscribers: Available at full price

Timeline: Launch end of Feb/ beginning of March

---

2. RAG Continuation: Graph DBs + ColPali (Advanced Phase)

As Promised, We’re Going Deeper

Phase 1 built a production-grade RAG system and built the Agentic Rag (advance part). Now we’re extending it - because no one builds production RAG in one shot. Real systems evolve as you discover what breaks and what users actually need.

Why These Extensions?

1. Evaluation (RAGAS, LLM-as-judge): You built RAG. But is it good? Measure faithfulness, relevance, and quality systematically instead of guessing.

2. Graph RAG: Vector search finds similar chunks. Graph RAG connects information across documents through relationships - answering multi-hop questions like “Who founded a startup after working at Google?” when the info spans multiple papers.

3. ColPali/VLMs: Your PDFs have tables, charts, complex layouts. OCR i.e. Docling breaks on these. ColPali treats pages as images, understanding both text AND visual elements without losing context.

   

This Is How Production Systems Actually Evolve

Build basic RAG → Measure it → Add relationship understanding →
Handle multimodal content.

You don’t jump to VLMs on day one.
You iterate based on what’s failing.
That’s production reality.

What You’ll Build:

Extending the existing arXiv RAG system:

• Evaluation pipelines with RAGAS and LLM-as-judge

• Graph RAG with Neo4j for multi-hop queries

• ColPali/VLM for visual document retrieval

Format:

• Paid subscribers only

• Detailed blogs + deep-dive videos

• Code extends arxiv-paper-curator repo

Timeline: Deliveries will be spread out

---

3. NLP Course: Zero to Fine-tuning

A New Addition to Our Roadmap - And Here’s Why

This wasn’t part of the original “The Mother of AI project”, but we realised it’s critically important.

The Problem We are Solving

Everyone’s jumping straight to LLMs and fine-tuning without understanding the fundamentals. We see people using GPT for tasks that TF-IDF or a fine-tuned BERT model could handle at 1/100th the cost. This isn’t just inefficient - it’s setting people up to fail when they hit production constraints.

Why This Course Matters

Shantanu has been working in NLP for almost 8 years, starting from absolute scratch with the basics. He’s seen every wave - from traditional NLP to word embeddings to transformers to LLMs. And here’s what’s being lost: the traditional approaches that still work brilliantly for many production use cases.

The gap in the market is real. Nobody talks about when TF-IDF outperforms embeddings. Nobody teaches you why a fine-tuned RoBERTa might be better than calling an LLM API. Everyone’s chasing the latest model without understanding the fundamentals.

What You’ll Learn

We’re covering everything important in NLP - from relevant basics to advanced techniques that are still critical today:

• Traditional NLP - TF-IDF, word2vec, text processing fundamentals

• Deep Learning for NLP - RNNs, LSTMs, seq2seq models

• Transformers Era - BERT, RoBERTa, understanding attention mechanisms

• Fine-tuning - Both older models (BERT, RoBERTa) and modern SLMs (Small Language Models)

• Production Decisions - When to use traditional NLP vs fine-tuning vs LLMs (cost, latency, accuracy trade-offs)

This is the progression every NLP practitioner should understand before jumping to LLMs.

Community-Driven Approach

This course will be open source and built together with the community. We can’t cover every NLP use case alone, so we’re building this collaboratively:

• All code open-sourced on GitHub, properly organized and documented

• Community contributions welcome (use cases, examples, improvements)

• Shantanu will guide the structure and quality, but we’re building this together

• More details on contribution process to follow

Format:

• Free for everyone (subscribers and non-subscribers)

• Comprehensive blogs with hands-on exercises

• Open-source code on GitHub

• Videos for complex topics when needed

Timeline: To be announced (we’re planning the scope carefully to do this right)

---

4. Recommendation Systems with MLOps

This Is Our Deepest Production Expertise

If you want to learn RecSys from people who’ve actually built these systems at scale, you’re in the right place.

Why Us?

Shantanu: 6+ years building RecSys in banking, then led a team of 10 engineers to build a complete jobs recommendation system from scratch. That project’s massive success was the biggest driver of his career progression. Over 4 years of continuous advancement - deeply involved in data, models, low-latency serving, infrastructure, evaluation, and complete MLOps setup on AWS. Multi-stage architecture with specialized models serving millions of users with millions of daily recommendations.

The reality check? He’s spent $3000+ on RecSys courses. Even expensive “expert” courses were missing critical production knowledge that only came from building these systems in the real world.

Shirin: Built news recommendation for Germany’s biggest news company - where algorithms competed directly with human curators. Unique constraints, fascinating problems, super low latency with Spark, serving millions of users. A completely different problem space that complements Shantanu’s experience.

The Gap in the Market

Multi-stage RecSys (candidate generation → ranking → re-ranking) has been standard since 5+ years ago. Everyone teaches the architecture.

But here’s what nobody tells you: what happens INSIDE those stages is where production RecSys actually lives. The approaches vary massively. Spotify ≠ H&M ≠ jobs ≠ news. Same architecture, completely different implementations.

Most courses focus on the easy part (the architecture). We’re teaching the hard part (what actually works in production).

What You’ll Build

Everything.
We guarantee this will be a RecSys course like nowhere else.

Complete end-to-end system covering:

• Multi-stage architecture with production-grade implementations

• Data pipelines, model training, low-latency serving

• Cold start, real-time vs batch, A/B testing, metrics

• Full MLOps integration - feature stores, model registries, CI/CD, monitoring, deployment pipelines

We have built these systems from scratch and touched every part. This level of end-to-end experience is rare. The knowledge from it is even rarer.

MLOps is a huge part of production RecSys, so we’re pairing RecSys with MLOps. These systems take years to develop in production. We’re compressing that knowledge into a structured path.

This combines Phase 3 + Phase 4 of the Mother of AI Project announced earlier.

Format:

To be decided (cohort vs self-paced vs weekly series)

Timeline: Likely after NLP course

---

2026 Launch Timeline

Here’s when everything is launching:

Q1 2026 (Jan-Mar) - STARTING NOW:

• ✅ Friday Discussion Sessions (First one this Friday!)

• ✅ Monthly Paid Subscriber calls (RecSys walkthrough this Sunday!)

Q2 2026 (Apr-Jun):

• AI Agents Course launches end of February/beginning of March (self-paced with live Q&A)

• RAG Continuation begins (deliveries spread out)

Q2-Q3 2026:

• NLP Course: Zero to Fine-tuning (free, community-driven, timeline TBD)

Q3-Q4 2026:

• RecSys with MLOps (format and exact timing TBD)

We’re intentionally spacing these out. We’re not doing 3-4 shallow projects. We’re going deep, one at a time, replicating how production systems actually evolve. Quality over quantity.

---

What You Get: Free vs Paid Breakdown

We know access varies by project, so here’s the clear breakdown:

Free Subscribers Get:

• Weekly content: career tips, production experience blogs, AI/ML insights

• Most Friday Discussion Sessions (3 PM CET, live on Substack)

• NLP Course: Zero to Fine-tuning - complete access (blogs + code + videos)

• Project announcements and selected technical content

• Access to open-source code repositories

Paid Subscribers (Monthly + Annual) Get Everything Above PLUS:

• RAG Continuation: Graph DBs + ColPali - full access (blogs + videos + code)

• Monthly Deep-Dive Calls - technical walkthroughs of production systems

• Priority support in Q&A sessions

• Early access to new projects and content

Annual Paid Subscribers Get Additional Benefits:

• AI Agents Course at 50% discount (detailed blogs + deep-dive videos + live Q&A)

• RecSys with MLOps (format and exact discount to be decided)

• Additional discounts on future cohort-based courses

• Exclusive access to specific advanced technical content

---

Ready to Build Production AI Systems?

2026 is about depth over breadth. Quality over quantity. Real systems over demos.

We’re not chasing every trend. We’re teaching what actually works in production - the systems we’ve built, the mistakes we’ve made, the lessons we’ve learned serving millions of users.

Start This Week:

• This Friday (3 PM CET): Join our first Discussion Session on Data Science Roadmap

• This Sunday: Paid subscribers - RecSys Systems Walkthrough

• Coming end of Feb/beginning of March: AI Agents course launches

Whether you’re here to level up your career, build production systems, or finally understand how AI/ML works beyond tutorials - you’re in the right place.

Subscribe to stay updated. Upgrade to paid to go deeper. Join the community to learn together.

Let’s build 2026 together. 🚀