March 2026

There’s a post-mortem template that’s become so familiar in Web3 it’s almost comforting. Bear market. Liquidity dried up. The team is moving on to pursue other opportunities. Thank you for being part of this journey. It’s usually true, in the way that “heart failure” is a true cause of death. It tells you nothing about why. After watching a lot of Web3 projects come apart, some slowly and some overnight, I’ve landed on a cause that almost never […]

Ver mais

Vibe coding is about momentum. You open your editor, prompt an AI, stitch pieces together, and suddenly you have something that works. Maybe it’s messy. Maybe the architecture is not perfect. But it’s liveand working, and that’s the point. Then comes deployment. This is where the vibe usually dies. Suddenly, you’re reading about containers, load balancers, CI/CD pipelines, infrastructure diagrams, and networking concepts you never asked for. You wanted to ship a thing. Instead, you’re learning accidental DevOps. […]

Ver mais

How I designed a single Python backend that serves a real estate agency in Dubai, a dental practice in Brazil, and a food retailer — each with fully isolated AI behavior, context memory, and business logic — without any off-the-shelf automation tools. Most WhatsApp AI tutorials show you how to build a bot for one use case. You connect to the API, craft a system prompt, call OpenAI, send a reply. It works. But it doesn’t scale. When I started building AI automation […]

Ver mais

A DevSecOps engineer’s honest account of building an AI agent grounded in real tenant data — including the parts that didn’t work. A note on scope: This is an experimental project built with a vision for future integration into a real business process. It is not yet production-grade. Enterprise AI agent platforms like Azure AI Foundry, AWS Bedrock Agents, and Google Vertex AI Agent Builder offer managed infrastructure, safety controls, and scalability that this implementation doesn’t. This is a learning exercise — with […]

Ver mais

arXiv:2602.23116v2 Announce Type: replace-cross Abstract: We consider the problem of contextual online RLHF with general preferences, where the goal is to identify the Nash Equilibrium. We adopt the Generalized Bilinear Preference Model (GBPM) to capture potentially intransitive preferences via low-rank, skew-symmetric matrices. We investigate general preference learning with any strongly convex regularizer and regularization strength $eta^{-1}$, generalizing beyond prior work limited to reverse KL-regularization. Central to our analysis is proving that the dual gap of the greedy policy […]

Ver mais

arXiv:2602.20293v2 Announce Type: replace-cross Abstract: We study a discrete denoising diffusion framework that integrates a sample-efficient estimator of single-site conditionals with round-robin noising and denoising dynamics for generative modeling over discrete state spaces. Rather than approximating a discrete analog of a score function, our formulation treats single-site conditional probabilities as the fundamental objects that parameterize the reverse diffusion process. We employ a sample-efficient method known as Neural Interaction Screening Estimator (NeurISE) to estimate these conditionals in the diffusion […]

Ver mais

arXiv:2602.06775v2 Announce Type: replace-cross Abstract: We study the problem of learning robust classifiers where the classifier will receive a perturbed input. Unlike robust PAC learning studied in prior work, here the clean data and its label are also adversarially chosen. We formulate this setting as an online learning problem and consider both the realizable and agnostic learnability of hypothesis classes. We define a new dimension of classes and show it controls the mistake bounds in the realizable setting […]

Ver mais

arXiv:2512.23075v4 Announce Type: replace-cross Abstract: Policy gradient methods for Large Language Models optimize a policy $pi_theta$ via a surrogate objective computed from samples of a rollout policy $pi_{text{roll}}$. However, modern LLM-RL pipelines suffer from unavoidable implementation divergences — backend discrepancies, Mixture-of-Experts routing discontinuities, and distributed training staleness — causing off-policy mismatch ($pi_{text{roll}} neq pi_theta$) and approximation errors between the surrogate and the true objective. We demonstrate that classical trust region bounds on this error scale as $O(T^2)$ with […]

Ver mais

arXiv:2512.21411v2 Announce Type: replace-cross Abstract: Singular learning theory characterizes Bayesian models with non-identifiable parameterizations through two central quantities: the real log canonical threshold (RLCT), which governs marginal likelihood asymptotics, and the singular fluctuation, which determines second-order generalization behavior and the complexity term in WAIC. While the geometric meaning of the RLCT is well understood, the interpretation of singular fluctuation has remained comparatively opaque. We show that singular fluctuation admits a precise thermodynamic interpretation. Under a tempered (Gibbs) posterior, […]

Ver mais