As a successful serial B2B, B2B2C product manager & product owner (0-1, 1-1000) Kurt’s career as a digital executive spans healthcare, media, laboratory automation, product management & development, digital transformation, strategy and digital marketing. His first start-up, PowerOne Media, was focused on the cutting-edge online media space. At PowerOne Media, Kurt led the product teams and business development to a market leading position of 1,800 local media customers.

Kurt co-founded his latest venture, Somml Health, in 2017. Somml Health is a Digital Health Neighborhood that optimizes care delivery through connected data and communications between the patient, their family, and health care teams, in real-time.

Through our highly scalable platform, Somml creates quantifiable opportunities for better health and more efficiencies in care delivery by elevating personalized care, and connecting daily insights and family support to enable early intervention. We believe whole person health takes a community, connected.

Expertise in: Digital Health, Patient Engagement, Innovation, Strategy, Product Management, Project Management, Mobile, Digital Marketing, Audience Development, Content Marketing, Content Distribution, Analytics, Search, Digital Media, Monetization, Business Development

Kurt Lozier

I am a Founder by day and a Professor by night. My work revolves in the realm of NLP and Multi-Modal Systems. We created traversaal.ai with one vision: to provide scalable LLM Products for  Enterprises, which can seamlessly integrate within the existing ecosystem while being customizable and cost-efficient.

If you are building Gen AI capability within your infrastructure and finding it challenging, we might just have a solution for you that can help you go to market faster and tremendously reduce cost. 

For speaking, advisory, or collaboration inquiries, please send an email to: hello@traversaal.ai

Learn more about my courses on LLMs here: https://maven.com/boring-bot/

Background: I have 15+ years of experience of leading Data Science/ Machine Learning teams. My area of expertise is Experiment Design, NLP, Recommend-er Systems and Time Series forecasting.

My experience spans over three continent and seven countries across APAC, Middle East and Africa and North America and multiple industries, namely, Tech, Telecommunications, Finance and Retail.

Specialties: Large Language Models,Causal Inference, Deep Learning, Graphical Modeling, Cloud Computing, Retail Analytics,  Optimization, Game Theory, NLP

Hamza Farooq

Last week, Lenny Rachitsky released 300+ podcast transcripts. Most people rushed to build quick no-code chat interfaces. We took a different path.

Instead of building another no-code frontend, we spent 72 hours engineering a production-ready knowledge graph RAG system that can answer the toughest questions about product, growth, and leadership—with full grounding and source attribution.

The result?
A system that doesn't just retrieve text. It understands:
→ Entities (concepts, frameworks, people)
→ Relationships (how ideas connect)
→ Context (when and why things matter)

Ask it "What is the explore vs exploit framework for career decisions?" and it synthesizes insights from multiple conversations, shows you exactly where Ada Chen Rekhi discussed it, and connects it to related mental models.

We open-sourced everything.
✅ One-command setup (5 minutes)
✅ 300+ indexed transcripts (Lenny, Julie Zhuo, Shreyas Doshi, Brian Balfour, and more)
✅ Beautiful Streamlit UI + CLI
✅ Local Qdrant vector DB (no Docker)
✅ Parallel processing (5-10x faster indexing)
✅ Hybrid search (local + global + vector)
✅ Smart caching & resume capabilities

Built costs ~$7. Runs locally. Query costs <$0.01.

This is what production-ready RAG looks like, not a demo, not a prototype. A real system you can learn from, fork, and extend.

🔗 Grab it on GitHub → https://lnkd.in/deh4aAVW

P.S. If you want to learn how to build AI systems like this from scratch, I'm teaching the Agent Engineering Bootcamp with $200 off (link in repo). But the code is free

🌍 230 million Urdu speakers. Zero standardized LLM benchmarks. Today we are releasing Urdu Bench (معیار) - world's first and most comprehensive benchmark for Urdu.

Most AI benchmarks focus on English, Chinese, and a handful of European languages. Urdu, spoken by over 230M people, has largely been invisible. 

With the launch of Alif in 2025, world's first bilingual Urdu-English LLM and recent advances by Taimoor Hassan (Qalb) and UpliftAI and many more amazing initiatives, we felt it was time to build a standardized benchmark.

So we built UrduBench, the first standardized (معیار) reasoning benchmark for Urdu that preserves context, logic, and symbolic structure.

🔍 What makes UrduBench different?
Most “multilingual” benchmarks rely on naive machine translation and call it done. We didn’t.

We used contextually ensembled translations with human-in-the-loop validation, ensuring:
✅ Context is preserved end-to-end
✅ Mathematical & logical structure remains intact
✅ Cultural nuances are validated by native speakers
✅ No machine-translation artifacts

📦 What’s included:
MGSM (math reasoning) — Urdu
MATH-500 (advanced problem solving) — Urdu
CommonSenseQA — Urdu
OpenBookQA — Urdu

You can’t claim “multilingual AI” without معیار.
And you definitely can’t claim it when 230 million speakers are missing from your evaluation pipeline.

UrduBench enables معیاری evaluation across:
- Diverse reasoning tasks
- Multiple difficulty levels
- Different model scales
- Cross-language consistency

🔜 What’s next:
Integration into the LM Evaluation Harness, making معیار-driven Urdu evaluation a default—not an afterthought.

🔗 UrduBench → https://lnkd.in/gdd2aDUT

A big congratulation to our Traversaal.ai team for such an amazing release!

Ali Shafique Shaza Khawaja

If you want to join us and work on similar research, please email us: jobs@traversaal.ai

I’m happy to share that I’m starting a new position as Instructor at Udemy! Starting the new year with the commitment to democratize AI reach for everyone!

Most AI agents are “clever parrots”-- They talk, but they don’t think.

Without memory, intelligence collapses into AI slop. It responds reactively, but forgets everything the moment the conversation ends. While it might work well enough for trivia, it’s ultimately useless for enterprise workflows. 

If you want personalized, adaptive AI agents, you need memory. 

I define four layers of agent memory that mirror human cognition:

Working Memory (Short-Term): Holds recent context, tool outputs, or reasoning chains. Like mental math, but fast and temporary.

Episodic Memory (Long-Term): A journal of past interactions. Critical for customer service agents that must recall user history or preferences.

Semantic Memory (Long-Term): Encodes facts and concepts. You don’t remember when you learned gravity, but you know it.

Procedural Memory (Long-Term): The agent’s muscle memory. How to call APIs, run tools, or execute SQL queries.

By building these specialized digital workers that don’t hallucinate or forget the information you fed it an hour ago, they can master domains like medical coding or supply chain forecasting to deliver repeatable, enterprise-grade results. 

Interested in utilizing AI to its fullest extent? I’m teaching a Maven course on building and deploying actual AI agents that drive real business impact. You’ll learn how to:

Prototype fast with no-code tools
Scale complex systems with low-code methods
Build full-scale web apps with AI backends

Join my next cohort in January with this exclusive offer for my LinkedIn audience: https://bit.ly/45bWA4P

🔥 Hot take: If your RAG breaks every 3 months, you don’t have a RAG problem, you have a fundamentals problem. This is exactly what is addressed in my book's latest chapter.

We’ve normalized building RAG on top of endless frameworks, only to:
-Fight dependency hell
-Rebuild pipelines after every minor update
-Debug abstractions we don’t actually understand

And then we wonder why our “production RAG” never survives.

That’s exactly why Chapter 6 of my book is all about building Retrieval-Augmented Generation (RAG) from scratch (team Manning Publications Co., thank you everything you do!)

No magic wrappers. No brittle glue code.

Just:
-Encoders + decoders working together
-Chunking that actually makes sense
-Vector databases used intentionally (yes, beyond FAISS)
-And real RAG evaluations, not vibes-based demos

I’m also excited to include a guest contribution from Andre Zayarni, co-founder of Qdrant — my favorite vector DB (non-sponsored 🙂) sharing insights from the infrastructure side that most tutorials completely ignore.

And here’s the kicker 👇
RAG is not the endgame.

📌 Chapter 7 is where things level up: Agentic RAG.
RAG systems that:
-Decide when to retrieve
-Decide what to retrieve
-Iterate, reason, verify, and improve answers autonomously

👉 Get the book here:
https://lnkd.in/d7prAqis

Frameworks come and go.
First principles compound.

🚀 RAG vs Knowledge Graphs with built-in Evals - let's settle this battle once and for all! 

Everyone argues about RAG vs Knowledge Graphs. Nobody actually measures which one wins. So I built a system that settles the debate with data. RAG entirely from scratch, and KGs using Neo4j

Here's the setup:
Same question. Two completely different approaches. One LLM judge scoring them head-to-head.

→ RAG uses semantic search + GPT synthesis → Knowledge Graph converts questions to Cypher on Neo4j → GPT-4o-mini evaluates both objectively

The results:
Knowledge Graphs crush on:
"Who are Emily Chen's collaborators?" (100% precision)
"How many articles published?" (exact counts)
"Which researchers work on AI Ethics?" (structured filtering)

RAG destroys on:
"What are challenges in AI safety?" (nuanced synthesis)
"Explain transformer innovations" (contextual understanding)
"Summarize ethical concerns" (semantic reasoning)

Neither is "better." They're built for different jobs.

What I open-sourced:
✅ Complete RAG vs KG comparison system 
✅ LLM-based objective scoring 
✅ Interactive graph visualizations 
✅ Query-specific subgraph rendering 
✅ Text-to-Cypher generation (no hardcoded queries, built from scratch)

Cost: ~$1 for full evaluation using Neo4j (free tier) + OpenAI.

This is what production AI looks like: knowing when to use structured queries vs semantic search, backed by objective data.

Want to learn how to build evaluation frameworks and hybrid AI systems like this? 

I teach the complete architecture in my Agent Engineering Bootcamp: RAG, Knowledge Graphs, multi-agent orchestration, and production patterns.

My next cohort kicks off February 4. Join us with this LinkedIn-exclusive offer: https://bit.ly/4qutaay

Github Repo: https://lnkd.in/gVJYB4Ws

I can't believe my book is Deal of the Day today on Manning Publications Co. Think of it as Best Seller for a Day 😂

Save 45% on my book Build an LLM Application (From Scratch) and other selected titles @ManningBooks: hubs.la/Q03-d27Y0

My experimental Airbnb project with enhanced search has now crossed 10k users! And best of all this was built without writing a single line of code!

For a simple query like:
“Find me a cozy 2-bedroom in Barcelona, under €100/night, walkable to Park Güell, with great WiFi for remote work, from 1st through 3rd Feb”

This tools just works beautifully. Best part: Totally built on no-code platforms!

No dropdowns. No checkbox gymnastics. Just intent.

This agent:
-Understands natural language
-Searches listings intelligently
-Personalizes results and explains why something is a good fit

This is how users actually want to search.

If you want to learn how to build agents like this — for real business workflows, not demos — I teach a course on Building GenAI Agents for Enterprise.
✅ Build a production-ready agent in 7 weeks
✅ Design RAG systems that actually work
✅ Ship safely, at enterprise scale

🚀 Next cohort starts Jan 26

👉 Join here: https://bit.ly/47KRBZ0

airbnb project link in comments

Your AI agent works… until it doesn’t.
And by the time you notice, it’s already in front of users.

Most teams spend weeks building agents — prompts, tools, workflows — and almost zero time thinking about how to evaluate them. The result? Demos that look great, but systems you can’t trust, debug, or safely ship.

That’s why Gabriela de Queiroz and I are hosting a one-hour practical session on evaluation strategies for AI agents from what we've learned working with Fortune 100 and startups, across various industries.

We’ll focus on what actually matters in the real world:
- How to tell when an agent is working vs. just sounding confident
- How to design evals that catch failures before users do
- How evals turn agent behavior into something measurable, improvable, and production-ready

If you’re building (or planning to build) agents for real users, this is the missing layer between “cool prototype” and “reliable product.”

👉 Join here: https://lnkd.in/gz5T-y-k

Build agents you can trust.

Everyone asks how to build AI systems. Almost no one asks how to know they’re working.

That’s the gap most LLM demos fall into.

Evals aren’t a “nice to have” — they’re the line between:
- AI that sounds impressive
- and AI you can actually ship, trust, and scale

In practice, this means going beyond model fluency and measuring:
-Can it reason correctly?
-Can it retrieve the right information?
-Is it grounded — or hallucinating?
-Can an agent complete a full task and recover when things break?

Until you can answer those, you don’t have a system — you have a prototype.
If you want to learn how to design and evaluate AI agents that survive production, I am teaching a free class next week with Gabriela de Queiroz

👉 Join here: https://lnkd.in/gz5T-y-k

taplio

The Battle

Kurt Lozier

Hamza Farooq

Followers

Posts

Engagements

Likes

Performance ScoreHamza Farooq's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Kurt Lozier

Hamza Farooq

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Hamza Farooq's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreHamza Farooq's LinkedIn Performance