PAOLO’S BLOG / 23 ARTICLES & COUNTING
AI. Data. Graphs.
Ideas in practice.
I write about the systems I build, the research I’m exploring, and the engineering decisions behind useful AI. Read my articles on GraphRAG, NLP, agents, and production machine learning.
Timestral by Mistral AI: How 209M-Parameter Time-Series Foundation Models Just Killed Bespoke Forecasting Pipelines
On October 15, Mistral AI stepped outside language models and dropped Timestral, a family of open-weight Time-Series Foundation Models (5M to 209M parameters) built on Decoupled Horizon Forecasting. Timestral-209M ranks #1 across GIFT-Eval, TIME, and fev-bench while running on standard CPUs in 140ms. Here is the architectural deep dive, why it pairs with System-1 decision models like JEV, and how to deploy zero-shot probabilistic forecasting in production.
14 min readJEV by TypeSafe AI: Why the Future of Agentic Routing Belongs to System-1 Structured Decision Models
TypeSafe AI just launched JEV, a non-generative 'System One' model that is 200x faster and 450x cheaper than traditional LLMs. It recently helped an AI agent beat Pokémon Red. Here is the architectural deep dive on why structured decision models are the missing piece for production-grade autonomous agents and machine-to-machine pipelines.
6 min readOpenAI o1 and the 30-Second Reasoning Cliff: How to Architect Production Systems Around Test-Time Compute, Hidden Tokens, and Cognitive Routing
OpenAI o1 dropped on September 12, trading intuitive forward passes for hidden chain-of-thought search. But in production, test-time compute introduces a 15–45s latency cliff, 4,000-token hidden billing overheads, and severe inverse-scaling failure modes. Here is the architectural deep dive, the benchmark truth, and the production-grade Cognitive Routing Gateway I deploy to harness System-2 reasoning without blowing enterprise budgets.
17 min readDeepSeek V4 Flash Just Made Your AI Budget Obsolete: The 0731 Release, the Architecture, and the Production Routing Strategy That Cuts Costs by 90%
DeepSeek V4 Flash 0731 dropped on July 31 with 284B parameters, 13B active, and $0.14/M input tokens, outperforming models that cost 100x more on agentic benchmarks. Here is the architecture breakdown, the benchmark reality, and the exact routing strategy I deploy in production to use it alongside Claude and GPT without compromising quality.
13 min readYour AI Agent Is a Liability: The Production Checklist That Separates $500/Day Contractors from $2,000/Day Architects
68% of companies blew their AI agent budget this year. 40% of agentic projects will be cancelled by 2027. The difference between agents that ship and agents that sink is not the model, it is the 14-point production checklist that most teams skip. Here it is, with the architecture I deploy for every engagement.
16 min readGraph Algorithms Your Enterprise Is Not Using: PageRank, Community Detection, and Centrality for Fraud, Churn, and Supply Chain
Most companies use graph databases as glorified foreign key stores. They store relationships but never analyze them. Here are five production-ready graph algorithms, each mapped to a concrete business outcome, with Python and Neo4j code you can deploy this week.
16 min readThe 94% Problem: Why Almost Every Company Deploying AI Is Failing, And the Exact Playbook That Works
Only 6% of enterprises capture significant value from AI. The rest are burning budget on pilots that never ship. After dozens of engagements, here is the post-mortem, and the five-step playbook the 6% actually follow.
14 min readGraph Foundation Models Are Here: The GPT Moment for Relational Data Has Arrived
LLMs learned language. Diffusion models learned images. Graph Foundation Models are learning relationships, and they are about to make every fraud detector, recommendation engine, and drug discovery pipeline you have obsolete. Here is what changed, why it matters, and how to build with GFMs in production.
14 min readStop Building Chatbots: Why 'Agentic AI' is the Only Data Science Trend That Matters in 2026
Conversational AI is dead. Autonomous AI agents are taking over enterprise workflows in 2026. Here is why you need to transition from building chatbots to deploying agentic swarms.
4 min readThe Knowledge Graph Stack for Agentic AI: LazyGraphRAG, Temporal Graphs, and the Architecture That Replaces Vector-Only RAG in 2026
Vector RAG was the 2024 starter kit. In 2026, production agents run on temporal knowledge graphs with skeleton indexing, DRIFT search, self-healing edges, and A2A coordination. Here is the new stack, with Python, Neo4j, and the cost math that makes it real.
21 min readMCP in Production: The Practical Guide to Model Context Protocol for AI Agents That Actually Ship
MCP is the protocol every AI agent now speaks. But shipping it in production, with auth, sandboxing, and tool governance, is where most teams stall. A hands-on architecture guide with Python, security patterns, and the cost math that makes it work for B2B.
16 min readBeyond RAG: Implementing Karpathy's 'LLM Wiki' Pattern with Obsidian
Why standard vector RAG fails at complex synthesis, and how to build a stateful, asynchronous LLM-maintained knowledge base using Karpathy's LLM Wiki pattern.
10 min readContext Engineering Is the New ETL: Why Prompt Engineering Died in 2026
Prompt engineering was a 2023 parlor trick. Context engineering is the discipline winning enterprise AI teams are building in 2026, a pipeline problem that looks more like ETL than copywriting. Here is why it matters, how it works, and where the ROI lives.
15 min readThe AI Bubble and the Return to Reality: Why an XGBoost Model Generates More ROI Than Any LLM
Most enterprises are burning budget on generative AI proofs of concept that solve problems nobody has. Meanwhile, the unsexy predictive models on tabular data, churn, pricing, fraud, keep quietly delivering 10–50x ROI. Time for a correction.
8 min readCompound AI Architecture: Semantic Routing for 90% Inference Cost Reduction
How to build a three-tier inference stack that routes user queries between local SLMs and frontier APIs using semantic classification, with Python implementation, Neo4j-backed domain routing, and a realistic cost model for B2B workloads.
13 min readThe SLM Flippening: Why 1B–8B Parameter Models Are Winning the Enterprise and the Edge
Why autoregressive inference is memory-bandwidth bound, how SLMs exploit distillation, GQA, and aggressive quantization, and how production stacks route cheap local models before touching frontier APIs.
7 min readBeyond 4-bit: How BitNet.cpp and 1.58-bit Ternary LLMs are Making GPUs Obsolete for Inference
Multiply-accumulate limits in FP Transformers, why PTQ keeps you memory-bandwidth bound, the ternary BitNet b1.58 formulation with AbsMean activations, and how BitNet.cpp maps inference to SIMD-heavy CPU kernels.
7 min readDeepSeek V4 and the mHC Breakthrough: How Manifold-Constrained Hyper-Connections Stabilize Trillion-Parameter Training
A deep technical dive into Manifold-Constrained Hyper-Connections (mHC), the architecture behind DeepSeek V4's stable training at 1T+ parameters. Birkhoff polytopes, doubly stochastic matrices, and why this changes the economics of large-scale AI.
8 min readFine-Tuning Qwen2.5 Locally: My Experience on an RTX 4060
A hands-on guide to fine-tuning Qwen2.5 on consumer hardware, LoRA, 4-bit quantization, and practical lessons from training on an NVIDIA 4060.
5 min readDeep Agents: From BabyAGI to Production-Ready Autonomous Systems
How task-driven autonomous agents like BabyAGI work, why they matter for complex workflows, and what it takes to move from prototype to production.
4 min readOpenClaw: The Open-Source AI Assistant That Actually Does Things
A deep dive into OpenClaw, the privacy-first AI assistant that runs locally and automates your digital life across messaging apps, email, and more.
5 min readBuilding a Customer Churn Model: From Raw Data to Production
A complete walkthrough of developing and deploying a churn prediction model, data prep, training, tuning, evaluation, and production deployment.
6 min readHello GraphRAG: Building a Graph-Aware RAG Pipeline with Neo4j
A step-by-step walkthrough of wiring a minimal GraphRAG prototype in Python using Neo4j and an LLM.
2 min read