Digital Factory AI Lead

Reshaping industry workflows by developing AI solutions and effective agentic systems.

Lead AI Engineer

moatless

Helping industry experts and researchers to navigate the sheer endless sea of Large Language Model (LLM) research & innovation because everybody deserves to find the best information.

Founder

Stealth

I was part of a 6 month government-funded Entrepreneurship program that enabled me to work on my own projects and ideas within the LLM space.

During this time, I gained experience in building LLM-powered backends with popular in-demand tools. The prototype I ended up building included the following:

- Fetching data from an external API & processing the response (Python)
- NoSQL database (MongoDB, Postgres)
- PDF Processing (PyPDF2, PyMuPDF)
- Document Extraction (Textract/Document Transformer)
- LLM Chain with multiple steps and prompt templates (Python, Guidance/LangChain)
- API endpoint for the processed data & the LLM outputs (FastAPI)
- Cloud Infrastruture (AWS)

Since I operated on a minimal budget, efficiency and cost reduction were a priority for me. Specifically, I focused on optimizing tasks like PDF processing, document understanding, information extraction and summarization.

Senior Machine Learning Engineer

A1 Telekom Austria AG

During my time as a Senior at A1, I was the DS lead for the company's most advanced NLP projects.

My work included

- Reporting results to several levels of management 
- Representing cutting edge Deep Learning & NLP at the company
- Pitching these topics to the C-Level and other influential exexutives
- Giving expert talks & presentations
- Participating in workshops
- Taking part in the recruiting process for full-time Data Science internships
- Guiding selected interns and assigning them roles in my projects

For a technical overview of my work, see below.

Senior Data Scientist

I was in charge of a Social Media Listening project for Facebook, Twitter and Online Newspapers which included the following

- SQL & MongoDB databases
- Web scraper for the newspaper(s)
- Processing Facebook & Twitter API data
- Transfer Learning & Language Model Finetuning on in-domain texts
- Multiclass and Multilabel Text Classifiers based on Embeddings and Transformers (the original PyTorch implementation of BERT)
- Tableau Dashboard

On top of that, I was developing intents & entities for customer text channels, including e-mail, web search, specialized chatbots (e.g., billing), the aforementioned Social Media data and several smaller Text & Sequence Classification projects.

I then moved on to explore speech channels and became the person in charge of Speech-to-Text at the company, which included 

- Processing short IVR audio recordings
- Processing long(er) call center recordings 
- Labeling pipeline with custom Prodigy scripts
- Developing a subtitle matching algorithm for aligning audio with text & filtering out errorenous subtitles
- Developing Hybrid HMM-TDNN Speech Recognition models with Kaldi
- Switching to (back then freshly released) bleeding edge libraries & models for End-to-end Deep Learning, such as NVIDIA NeMo & Facebook FAIR
- Improving performance on Austrian dialects & in-domain recordings
- Adapting the in-house Classification models mentioned earlier for the text output of the speech channels

In addition, I was able to work on PoCs in the areas of Document Extraction, Document Classification, Entity Extraction & Question Answering.

Data Scientist

Universität Wien

Research in Cognitive & Economic Psychology:

- Dual process theory (Type I/II thinking)
- Metacognition theory
- Experimental assessment of cognitive biases
- Leveraging aggregated mouse cursor movement data

Junior Researcher

Coursera

Data Science Education

As the founder of LLM Watch, my goal is to curate the most relevant Large Language Model (LLM) research and make it accessible to a broad audience. 

I distill complex technical topics into summaries that respect your time. 

Before embarking on this journey, I led Deep Learning R&D at one of Europe's largest telecom companies. There, I spearheaded numerous AI projects with a focus on Natural Language Processing (NLP) and Automatic Speech Recognition (ASR). 

With more than 6 years of experience in AI, I was there before the hype. 

I try to offer a healthy balance between shiny new things and what really matters.

Pascal Biese

Yann LeCun's first big paper after he left Meta. Here's what they did:

For years, training Joint-Embedding Predictive Architectures (JEPAs) has been more art than science. Stop-gradients, teacher-student networks, carefully tuned EMA schedules – a delicate dance of heuristics to prevent representation collapse. But what if there was a principled way?

Self-supervised learning promised to unlock the power of unlabeled data, yet state-of-the-art methods like DINOv2 and I-JEPA rely on complex recipes with dozens of hyperparameters. Each new dataset or architecture requires extensive tuning. Even worse, the training loss often doesn't correlate with downstream performance, making model selection nearly impossible without expensive labeled evaluation.

The core problem: we've lacked a theoretical understanding of what makes good representations. LeJEPA changes this by proving a fundamental result – isotropic Gaussian embeddings uniquely minimize worst-case risk across downstream tasks. 

Tho achieve this, they introduce SIGReg (Sketched Isotropic Gaussian Regularization), a novel objective that enforces this optimal distribution using random projections and characteristic function matching. The beauty lies in its simplicity: linear complexity, bounded gradients, and provably correct by construction.

LeJEPA eliminates stop-gradients and teacher-student networks entirely while achieving competitive performance – 79% on ImageNet-1k with a frozen ViT-H/14. And here's the best part: it works out-of-the-box across 60+ architectures and diverse domains - in-domain pretraining with LeJEPA on small datasets (even just 1,000 samples) outperforms transfer learning from massive frontier models trained on billions of images.

This could fundamentally shift how we think about foundation models. Instead of relying on ever-larger generic models, practitioners in specialized domains can now train competitive models directly on their data with minimal hyperparameter tuning. 

The fact that LeJEPA's training loss actually correlates 99% with downstream performance also solves the critical model selection problem that has plagued SSL research.

What do you think? Is Yann onto something?
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

AI Lead at PwC <​/​> Daily AI highlights for 80k+ experts 📲🤗

LeJEPA: Provable and Scalable Self-Supervised Learning

Microsoft just announced the "Era of Agentic Organization". Here's what that means:

As LLMs scale their reasoning capabilities, we need new techniques: sequential thinking is too slow, and parallel thinking wastes compute by running independent traces that can't learn from each other mid-process. The industry needs a paradigm that combines the adaptivity of sequential reasoning with the speed of parallelization.

For this reason, Microsoft Research developed AsyncThink. AsyncThink is a new approach where a single LLM learns to play two roles simultaneously: an organizer that dynamically structures the reasoning process through Fork and Join operations, and workers that execute sub-queries concurrently. 

Their key contribution is that the model learns via reinforcement learning to optimize its own thinking structure - deciding when to split problems, which sub-queries to pursue, and when to merge intermediate results.

The results? 28% lower inference latency compared to parallel thinking while maintaining higher accuracy on mathematical reasoning. Models trained only on simple countdown problems successfully generalize their asynchronous thinking capabilities to completely unseen domains like Sudoku and complex math problems.

The more tasks AI is supposed to solve within an organization, the more important it becomes to structure all of this in a scalable and controllable way. AI agents need to learn to organize their reasoning like skilled project managers, adaptively coordinating multiple thinking processes based on the problem at hand. 
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

The Era of Agentic Organization

Can we train billion-parameter models without gradients? Oxford researchers just showed it's possible - at 100x the speed.

Evolution Strategies have always been an elegant alternative to backpropagation. They don't need differentiability, handle noisy objectives gracefully, and parallelize beautifully. But there's been a fundamental problem: they simply don't scale to modern model sizes.

University of Oxford's new EGGROLL algorithm mirrors what made LoRA successful for fine-tuning - instead of perturbing all parameters with full-rank matrices, use low-rank approximations. For each population member, sample matrices A and B to form a rank-r perturbation AB^T rather than a full matrix E. This cuts memory from mn to r(m+n) per layer and reduces forward pass compute proportionally.

What makes this work is subtle but powerful: while each perturbation is low-rank, the final parameter update averages across the entire population, yielding a full-rank update when the population is large enough. The researchers prove this approximation converges to full-rank ES at an O(1/r) rate - faster than typical central limit theorem rates due to the symmetry of the problem.

The results validate the approach across three domains. In pure integer pretraining of recurrent language models, EGGROLL enables stable training with population sizes exceeding 250,000 - orders of magnitude larger than prior work. For LLM reasoning tasks, it matches or exceeds GRPO while maximizing hardware utilization. And critically, it achieves near-inference throughput speeds, making gradient-free optimization genuinely practical at scale.

This opens real possibilities for training systems that can't use backpropagation: end-to-end neurosymbolic architectures, models with discrete components, or deployment scenarios where integer-only computation is essential for efficiency.
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

Evolution Strategies at the Hyperscale

We might be systematically underestimating LLMs.

We've been told that reinforcement learning is essential for frontier reasoning capabilities. Massive compute budgets, curated datasets, reward models - the whole apparatus. 

But new research from Harvard University challenges this expensive paradigm with a surprisingly simple question: can we just sample better from base models?

The paper introduces a training-free algorithm that achieves reasoning performance matching or exceeding GRPO (the standard RL method) across mathematics, coding, and science benchmarks. 

The key insight revolves around the "power distribution" - sampling from p^α instead of p. Unlike low-temperature sampling which greedily averages future token likelihoods, power sampling accounts for entire future completion paths. 

This seemingly subtle difference makes a big difference: it naturally upweights tokens with fewer but higher-likelihood futures, essentially planning ahead rather than getting trapped in dead-ends.

Using Metropolis-Hastings MCMC to approximate this distribution, the researchers show that Qwen2.5-Math-7B can reach 74.8% on MATH500 without any additional training - nearly matching the 78.5% from full RL posttraining. 

On out-of-domain tasks like HumanEval, it outperforms by up to 20 percentage points. Perhaps most remarkably, unlike RL which suffers from diversity collapse, this method maintains strong pass@k performance across multiple samples.

The takeaway is that base models may already contain the reasoning capabilities we're spending enormous resources to extract through RL. This would open up a new axis for inference-time compute scaling that's accessible to anyone with a base model - no training infrastructure, curated datasets, or verifiers required.

We're still learning how to properly query the "intelligence" that already exists in our models.
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

Your Base Model is Smarter Than You Think

Anthony Alcaraz

Amazon Web Services (AWS)

Driving growth in the startup ecosystem by leveraging AWS's cutting-edge AI/ML technologies. My team helps startups accelerate their go-to-market strategies and technological development.

Key focus areas:
- Developing AI/ML startups
- Guiding high-potential startups through AWS's AI/ML offerings
- Fostering partnerships within the EMEA tech community
- Aligning AWS's AI/ML solutions with startup needs

Passionate about empowering the next wave of AI-driven innovation. 

Let's connect and explore how AWS can supercharge your startup's AI/ML journey!

Senior AI/ML Strategist Startups & VC EMEA Strategic Pursuits

Fribl

Advisor

GenAI x HR 
Designed AI strategy
Lead engineering team 

With recruiting topping the list of concerns for many CEOs and companies spending upwards of $4,000 on average per hire, the need for innovation is clear. 

All too often, the arduous recruiting process leaves both employers and applicants frustrated after 42 days of effort with uncertainty if the right match was made...

At Fribl , we are leading the charge to transform this status quo, starting with reinventing the screening process using our proprietary GenAI technology enhanced by symbolic AI. 

This powerful combination doesn't just accelerate hiring, but actually personalizes the experience to precisely match candidates with roles that align to the company's unique culture - a monumental undertaking previously thought impossible.

As innovations like video interviews, CV parsers and AI recommendations shake up recruitment, we remain laser focused on four seismic shifts: Diversity, Advanced Interviewing, Data Analytics and AI Adoption. 

Our mission? To disrupt outdated screening tactics by simplifying the complex, multidimensional analysis of applicant credentials at scale to surface the ideal talents that personify an organization's ethos.

Chief AI/ML Product Officer & Partner

University of Oxford

Guest lecturer to Ajit Joakar Course : Artificial Intelligence: Generative AI, Cloud and MLOps (online)
Subject : the intersection of LLMs and graphs

AI/ML Guest Lecturer

Aldecis

My client portfolio was L’Oreal Finance Transformation 
AI for BI 

Management in AI Projects and Engineering Teams:
In my managerial role, I oversee AI projects and lead teams of engineers in crafting cutting-edge solutions. My focus is on driving strategic initiatives, ensuring projects are aligned with organizational goals, and fostering a culture of innovation and excellence. I take a hands-on approach to problem-solving and mentorship, ensuring that both projects and people reach their fullest potential.

Development in LLM/ML Ops:
My development expertise lies in the specialized domains of Large Language Models Operations (LLM Ops) and Machine Learning Operations (ML Ops). I am responsible for streamlining machine learning workflows from development to deployment. My work enables organizations to not only adopt but excel in leveraging machine learning technologies, facilitating continuous learning and adaptability across the board.

Marketing and Strategic Partnerships:
In the realm of marketing, I focus on building and nurturing strategic partnerships. My efforts are geared towards creating synergies with like-minded organizations and industry leaders. By forming valuable partnerships, I amplify brand visibility, foster innovation, and open doors to new business opportunities. The result is a win-win situation for all stakeholders involved, enhancing both market reach and impact.

Chief AI/ML Product Officer

CNAF

AI for performance management and customer support

I served as a Senior Executive after a competitive exam (EN3S). 
In this multifaceted role, I spearheaded a range of initiatives across key sectors such as Finance, Fraud Detection, Quality Control, Accounting, and Data & Performance Management.

Signature Project in ML Ops:
Facing the pervasive challenge of managing recurring incoming flows, I took the lead in conceptualizing and executing an avant-garde Proof of Concept (POC). This POC employed Azure ML-based ML Ops solutions, setting a precedent by being the organization's inaugural foray into Machine Learning Operations. The result was a transformative improvement in operational efficiency and service delivery.

Influence on Pluriannual Strategy 2023-2027 :
My pioneering project had a profound impact on the organization's multi-year strategic plan, specifically in the following domains:

Predictive Analytics: My efforts catalyzed the adoption of predictive analytics, enabling the organization to preemptively identify and mitigate challenges in benefits delivery. This proactive methodology revolutionized our service approach.

Sentiment Analysis: The launch of the "Lab Parentalité" was a critical milestone. Through the use of sentiment analysis, we gained unprecedented insights into parent feedback. This initiative not only enriched our data but also helped us better comprehend and meet the needs of parents.

Chatbots & Virtual Assistants: I played a key role in overhauling our customer service paradigm. By integrating cutting-edge chatbots and virtual assistants, we were able to enhance customer engagement across multiple platforms, including phone support, online portals, and mobile apps.

Personalization Algorithms: To elevate the user experience, we incorporated algorithms designed to customize services and recommendations. This bespoke approach significantly increased user satisfaction by tailoring the experience to individual preferences and needs.

Writes on ML/LLM Ops 🎨 | Follow me (4,8 K+) 🔔

As AWS's Senior AI/ML Strategist for EMEA Startups, I specialize in helping early-stage AI companies accelerate their development and go-to-market strategies. My expertise spans:

Technical Domains:
- Large Language Models (LLM) and GenAI architectures
- Knowledge Graphs and Vector Spaces
- MLOps and production AI systems
- Multi-agent and agentic AI systems

Strategic Value-Add:
- Technical due diligence for  AI investments
- Architecture optimization for AWS AI/ML services
- GTM strategy development for AI startups
- Deep expertise in evaluating GenAI applications

Recent Achievements:
- Guest lecturer at Oxford on LLMs and graph-based systems
- Led AI strategy for multiple venture-backed startups
- Published thought leadership on agentic AI systems
- Built and scaled ML platforms for enterprise clients

I'm passionate about helping VCs and founders navigate the technical and strategic challenges of building transformative AI companies. Happy to provide technical evaluation, architectural guidance, and strategic insights to help portfolio companies succeed.

Chief AI Officer & Decision scientist

Management in AI Projects and Engineering Teams:
In my managerial role, I oversee AI projects and lead teams of engineers in crafting cutting-edge solutions. My focus is on driving strategic initiatives, ensuring projects are aligned with organizational goals, and fostering a culture of innovation and excellence. I take a hands-on approach to problem-solving and mentorship, ensuring that both projects and people reach their fullest potential.

Development in LLM/ML Ops:
My development expertise lies in the specialized domains of Large Language Models Operations (LLM Ops) and Machine Learning Operations (ML Ops). I am responsible for streamlining machine learning workflows from development to deployment. My work enables organizations to not only adopt but excel in leveraging machine learning technologies, facilitating continuous learning and adaptability across the board.

Marketing and Strategic Partnerships:
In the realm of marketing, I focus on building and nurturing strategic partnerships. My efforts are geared towards creating synergies with like-minded organizations and industry leaders. By forming valuable partnerships, I amplify brand visibility, foster innovation, and open doors to new business opportunities. The result is a win-win situation for all stakeholders involved, enhancing both market reach and impact.

Your AI agents can't reason about what they can't structurally represent. ⭕ 

The gap between demos and deployable agents comes down to one thing: knowledge representation strategy.

Here's when knowledge graphs become non-negotiable for agentic systems:

🔗 Multi-hop inference is core to your reasoning tasks

Vector embeddings excel at semantic similarity but collapse on relationship queries. When your agent needs to traverse "Service → DEPENDS_ON → Database → PROTECTED_BY → SecurityGroup" to assess blast radius, you need graph topology. Property graphs give you O(1) relationship traversal. Document stores force you to reconstruct connections at query time.

⚡ You're hitting the relational complexity threshold

The 30-table heuristic is real. Beyond this, JOIN complexity explodes exponentially. More critically: when understanding connections requires traversing 3+ relationship types simultaneously, relational schema optimization becomes the bottleneck. Graph databases are optimized for exactly this access pattern.

🧠 Your agents must handle temporal reasoning and contradictory perspectives

Event-centric schema patterns capture causality. Multi-perspective patterns represent conflicting data sources with confidence scores and provenance. Try modeling "Terraform says X, AWS API says Y, confidence scores differ" in a relational schema without creating a maintenance nightmare.

🔄 Entity resolution across fragmented systems is blocking coherence

When the same entity appears as "Bob Jones" in CRM, "Robert Smith Jr" in fraud detection, and "R.J. Smith" in compliance, your agent needs evidence-based resolution using relationship patterns and contextual validation. Graph neighborhoods provide the evidence structure for identity inference.

⚙️ The three-graph architecture matters for production

Separate domain graph (trusted structured data), lexical graph (unstructured source text), and subject graph (LLM extractions). Entity resolution creates CORRESPONDS_TO edges between them. This prevents extraction artifacts from polluting canonical knowledge while maintaining full provenance.

❌ Skip graphs when:

Simple retrieval suffices, no multi-hop reasoning required, or relationship density is low. RAG with vector search handles these cases at 1/10th the infrastructure complexity.

The technical reality: RDF gives you formal reasoning semantics but slower performance. Property graphs give you traversal speed but require explicit inference logic. Hypergraphs handle n-ary relationships elegantly but lack mature tooling.

Choose based on your reasoning requirements, not architectural trends.

What's your biggest challenge in getting LLMs to reason over complex, interconnected enterprise data?

Senior AI/ML Strategist Startups @AWS | O’Reilly Author | Business Angel

Most agentic systems hardcode their capabilities. 🔳

This does not scale.

Ontologies as executable metadata for the four core agent capabilities can solve this.

🔍 OBSERVE: Annotation-Driven Discovery

Instead of hardcoding entity recognition logic, mark identifying properties with inverseFunctional annotations in your ontology.

The agent queries at runtime: "What property uniquely identifies this entity type?"

Result: Generic retrieval code that works across any domain. Your agent constructs type-safe queries dynamically based on schema metadata.

🧮 REASON: Declarative Logic Execution

Extend SWRL with three operational primitives:

Negation as Failure - handle incomplete information and default rules
Existentials - create derived entities during reasoning chains
Aggregates - compute across collections (sum, count, max)

Store business rules as graph data. The reasoner executes them. Every inference is traceable back through the rule chain.

⚡ ACT: Three-Tool Generic Architecture

Build these once, use everywhere:

Tool 1: Semantic discovery (vector search + ontology enrichment)
Tool 2: Dynamic retrieval (queries ontology for addressing scheme)
Tool 3: Context expansion (queries ontology for traversal rules)

Mark relationships with contextualizing annotations. The agent discovers which paths to follow by querying your schema.

📈 LEARN: Self-Referential Improvement

Here's where it gets interesting. The system queries its own structure.
"What entity types lack dedicated tools?"
"Which relationships have high traversal costs?"
"What query patterns appear frequently but aren't optimized?"

Track tool usage, success rates, execution times. Feed this back into ontology refinement.

The agent proposes new tool definitions based on usage gaps. It optimizes reasoning paths through technical rules that don't change semantics.

🏗️ The Engineering Stack:

Co-locate ontology nodes with data in your graph database
Use annotations as queryable metadata (not documentation)
Implement dynamic query construction from ontology responses
Build generic tools that operate on ontological primitives
Enable runtime tool discovery through schema queries

Your ontology is live, queryable metadata that directly influences execution. Not documentation. 

⚙️ Implementation Pattern:

Agent receives task
↓
Query ontology for relevant entity types
↓
Query ontology for identifying properties
↓
Query ontology for contextualizing relationships
↓
Construct and execute dynamic query
↓
Return results with provenance

🎯 Why This Architecture Wins:

Domain experts can extend capabilities by modifying the ontology. No code changes. No redeployment. The agent adapts immediately.

Business logic lives in declarative rules that outlive implementation technologies. Your ontology is a lasting asset, not code that rots.

Systems can analyze their own schema to identify improvement opportunities.

Meta-programming becomes possible.

Your LLM based agent can't reason about what it can't traverse. ❌ 

After analyzing production agentic systems at scale, I've identified why most AI agents fail: they're built on vector embeddings and relational schemas that fragment knowledge into isolated records.

The agents that actually work? They're built on graph-based knowledge models that encode relationships as first-class primitives.

🏗️ Architecture pattern that's winning: The three-graph model

Separate your knowledge graphs by epistemic confidence:

Domain Graph: Canonical entities post-entity-resolution. Stable identifiers, validated relationships, single source of truth.

Subject Graph: LLM extractions with explicit uncertainty. Confidence scores, model versions, extraction metadata preserved.

Lexical Graph: Original unstructured sources. Complete provenance, immutable text chunks, enables RAG retrieval.

The killer feature? CORRESPONDS_TO relationships link Subject → Domain only after entity resolution validates the match. Your agent knows what's certain vs. extracted.

⚙️ The technical tradeoff matrix I'm using:

Property Graphs (Neptune, Neo4j): Fast traversals, intuitive semantics, but inference requires explicit coding

RDF + OWL: Native reasoning via formal logic, but 3-5x performance hit and complexity overhead

Hypergraphs: Natural n-ary relations (prescription linking doctor, patient, medication, dosage simultaneously), but immature tooling

Most production systems? Hybrid LPG for performance + RDF layer for reasoning tasks.

🎯 Implementation pattern: Selective vectorization + ontology-driven traversal

Don't vectorize everything. Curate your vector search space:

Index: Service names, error signatures, entity anchors

Exclude: Full configs, log streams, operational noise

Vector search handles ambiguous queries. Ontological annotations drive precise graph traversal for contextualization.

Result: Semantic flexibility + structural reliability.

📐 Entity resolution is your foundation

Evidence-based matching with Jaro-Winkler similarity + context validation. Not LLM guessing.

Explainable: "92% match because name=87%, address=100%, phone=95%"

This generates the stable node identities everything else depends on.

The companies shipping reliable agents aren't prompt engineering harder. They're building knowledge architectures that make invalid reasoning paths literally untraversable.

I just published a deep dive on Agent-as-a-Service in XAnge's 2025 Seed Blueprint. ❕ 

After advising dozens of European AI startups at AWS, I'm seeing the same architectural mistake over and over.

Founders are bolting LLMs onto existing UIs and calling it "AI-powered." That's not innovation, that's technical debt with a chatbot.

Here's what actually works:

The Sidecar Pattern for Tool-ification

Deploy a lightweight MCP (Model Context Protocol) server alongside your app. Auto-expose your OpenAPI endpoints as agent-discoverable tools. Don't rip up your codebase, extend it.

Your "Search documents," "Create user," and "Generate report" functions become callable capabilities that any agent can orchestrate.

Why Reinforcement Learning Isn't Optional

Prompted agents are unreliable. Chain-of-thought reasoning breaks on multi-step workflows.

The path to production-grade agents: → Capture trajectories (action sequences + tool I/O) from day one → Build a semantic layer/knowledge graph for tool dependencies → Apply RL to train specialized models on your specific domain

Result? Smaller models that outperform GPT-4 on your task at 10x lower cost and latency.

The Data Architecture Most Founders Skip

You need three layers before your first agent ships:

Trajectory storage - every agent action, input, and output
Retrieval-augmented tool selection - serve only relevant tools per task (prevents prompt bloat)
Knowledge graph - model tool relationships deterministically, not probabilistically

It's the difference between a demo and a product.

Business Model Primitives

Usage-based pricing requires new infrastructure:

Authenticated requests (tamper-proof activity records)
Real-time credit ledger (transparency + compliance)
Automated payment rails (no service interruptions)

Enterprise customers won't adopt agents without audit trails and deterministic behavior. Build these primitives first.

The European Advantage

GDPR is a competitive moat. Data minimization, explicit consent, and transparent model behavior are trust primitives that shorten enterprise sales cycles.

Vertical industries (manufacturing, logistics, health) with structured workflows and compliance requirements are perfect for agent orchestration. Europe owns this domain depth.

What's Next

The agent marketplace is forming now. AWS launched ours.  Agentic browser for agent-discoverable services are emerging.
If your service isn't callable by other agents, with machine-readable schemas and semantic descriptions, you're invisible in this economy.

Teams that start this journey now with disciplined, data-first architecture will define the next decade of European software.

Are you building for reinforcement learning from day one, or retrofitting it later?

XAnge Seed Blueprint

AI agents are economically broken at scale without caching ❌ 

A single user goal triggers 5-10 LLM calls in loops. At 10K daily conversations, you're burning $547K/year on a single workflow pattern.

The obvious fix? Semantic caching with vector embeddings.

The problem? It fails catastrophically for agents.

🔴 Why Semantic Caching Breaks:

Agents execute context-dependent reasoning chains. Each step builds on previous steps.

"Analyze Tesla's Q3 performance" vs "Analyze Tesla's stock price"

Cosine distance: 0.28 → Cache hit 

Actual equivalence: False

Your agent retrieves a financial analysis plan when the user wanted technical specs. 

The entire workflow executes the wrong reasoning path.

Tune your threshold tighter → 5% hit rate (still burning $520K) Tune it looser → 65% hit rate but 40% false positives

The F1 optimization doesn't solve context drift.

🔧 The Architecture That Works:

Hybrid Cache = Vector Speed + Graph Intelligence

Layer 1: Vector DB for fuzzy matching (~30% of redundant queries) 

Layer 2: Knowledge Graph storing reasoning structures, not text
Explicit entity disambiguation (Apple fruit ≠ Apple stock)
Multi-hop traversal for validating context alignment 

Layer 3: Validation gate before serving cached plans

🎯 The Critical Shift: Granularity

Don't cache full responses. Cache atomic sub-queries.
"Compare platforms" decomposes into:
FetchFeatures(Platform A)
FetchPricing(Platform A)
FetchFeatures(Platform B)
CompareAttributes(A, B)

These canonical sub-tasks repeat at 40-70% rates vs 5-15% for full workflows.

The graph stores execution plan templates. 
Vector search finds intent. 

Graph traversal validates the reasoning path applies to current context.

📊 Production Results:

Hit rate: 65% (vs 8% with semantic-only)
Precision: 95% (vs 40% with loose thresholds)
Cost: $547K → $191K annually (65% reduction)
Latency: 60% faster (cold → hot state)

Graph adds 20ms per query, but prevents false positives that would cost 2000ms+ in LLM calls or failed workflows

The Mental Model:

Traditional cache: "Have I seen this text?" 
Semantic cache: "Have I seen something similar?" 
Agentic cache: "Do I know how to solve this class of problem?"

You're not storing what agents said. 
You're storing how agents think.

The cache evolves from a lookup table to a reasoning substrate, the agent's long-term memory of validated strategies.

Agentic needs graphs. 

At scale, the hybrid architecture : the difference between agents that burn capital and agents that generate ROI.

taplio

The Battle

Pascal Biese

Anthony Alcaraz

Followers

Posts

Engagements

Likes

Performance ScorePascal Biese's LinkedIn Performance

Performance ScoreAnthony Alcaraz's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Pascal Biese

Anthony Alcaraz

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);}Pascal Biese's LinkedIn Performance

Performance ScoreAnthony Alcaraz's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScorePascal Biese's LinkedIn Performance