Type to search posts and projects ↑↓ to navigate
Notes on building data platforms, AI systems, and the infrastructure between them. Technical deep-dives, lessons from the field, and occasional explorations of distributed systems architecture.
I wrote a fused Mixture-of-Experts dispatch kernel in pure Triton that beats Stanford's CUDA-optimized Megablocks at inference batch sizes, and runs on both NVIDIA and AMD GPUs without a single line of CUDA.
A loop always converges. That it reached a stable 'done' state tells you nothing about whether that state is correct, because the loop selects for whatever passes your check, not for what you meant. A practitioner's take on loop engineering: why convergence is not correctness, why the verifier is the part that actually decides, and what changes when loops run unattended over real production systems instead of a coding agent in a terminal.
Everyone assumes activation probes are too expensive to run in production. I ran the numbers on paper and I no longer believe it. Here is the argument, my predictions, and the harness I built to settle it.
Everyone assumes activation probes are too expensive to run in production. I ran the numbers on paper and I no longer believe it. Here is the argument, my predictions, and the harness I built to settle it.
What DeepSeek's DSpark and DeepSpec release actually changes for LLM inference: suffix decay, confidence-scheduled verification, a decision guide, and how to train a draft model for your own weights.
A loop always converges. That it reached a stable 'done' state tells you nothing about whether that state is correct, because the loop selects for whatever passes your check, not for what you meant. A practitioner's take on loop engineering: why convergence is not correctness, why the verifier is the part that actually decides, and what changes when loops run unattended over real production systems instead of a coding agent in a terminal.
Embedding space is the new ad real estate. Mapping LLM ad auctions, RAG poisoning, GEO, and a framework for what comes next.
I wrote a fused Mixture-of-Experts dispatch kernel in pure Triton that beats Stanford's CUDA-optimized Megablocks at inference batch sizes, and runs on both NVIDIA and AMD GPUs without a single line of CUDA.
Three labs. Three different bets on how to catch models misbehaving. Each makes different assumptions about when models 'know' they're lying. Here's what works, what doesn't, and what happens when you combine them.
OpenAI trained GPT-5 to confess when it misbehaves. It works surprisingly well - except when the model doesn't know it's misbehaving. That's where activation probes come in.
I've been working with steering vectors for months. Here's what actually works in practice, what fails in ways nobody warned me about, and the honest playbook for getting started.
A follow-up to my MCP Maturity Model post. Moltbook shows what happens when you run 770K agents at Level 0 maturity with zero governance. The results are instructive.
MIT Tech Review named mechanistic interpretability a 2026 Breakthrough Technology. Anthropic open-sourced circuit tracing. Here's what actually changed, how it connects to the activation probes I built for sandbagging detection, and why production teams should care.
Reinforcement Learning with Verifiable Rewards powers every reasoning model worth talking about. But it only works where you can check the answer automatically. Extending it to messy, real-world domains is the hardest open problem in LLM training right now.
MCP handles agent-to-tool. A2A handles agent-to-agent. A2UI handles agent-to-interface. Together they form a protocol stack that nobody has mapped properly - including the security gaps that should terrify you.
Interactive exploration of Manifold-Constrained Hyper-Connections - how DeepSeek fixed the signal explosion problem in deep residual networks using 1967 mathematics
First empirical demonstration of activation-level sandbagging detection. Linear probes achieve 90-96% accuracy across Mistral, Gemma, and Qwen models. Key finding - sandbagging representations are model-specific, and steering can reduce sandbagging by 20%.
I tested activation steering on 4 agent behaviors across 3 models. The results surprised me.
A deep dive into building distributed LLM evaluation infrastructure that actually scales - architectural decisions, trade-offs, and lessons learned.
A practical framework for evaluating your multi-agent context management strategy. From ad-hoc string concatenation to self-evolving context systems - where does your architecture stand?
Lessons from building a framework that automatically generates verified distributed systems - and why I think formal methods, synthesis, and ML need to work together
Enterprise data platforms face a 100,000x query increase from agentic AI. Introducing Symbiotic Agent-Ready Platforms (SARPs) - the architectural paradigm shift needed to survive the transition to machine intelligence.
Frontier AI models from OpenAI, Anthropic, Google & others can detect when they're being tested and modify behavior-challenging AI safety evaluation methods.
How the AI industry is responding to situational awareness challenges. Practical monitoring systems, collaborative research, and what organizations should do today.
Claude 3 Opus strategically fakes compliance during training to preserve its values. This alignment faking undermines our ability to modify AI behavior safely.
Researchers achieved a 30-fold reduction in AI scheming through deliberative alignment. But rare failures persist. Can we truly train models not to deceive?
Frontier AI models from OpenAI, Anthropic, and Google can now recognize when they're being tested. This observer effect undermines AI safety evaluation.
AI companies are getting sued over training data, agents operate with no permission framework, and users can't control their AI profiles. I wrote four open standards (LLMConsent) to create a decentralized consent protocol for AI - like HTTP but for data rights, agent permissions, and user sovereignty. This is an RFC, not a product.
Explore Kimi K2’s trillion-parameter MoE architecture, MuonClip optimizer, and agentic training. Learn why it outperforms GPT-4.1 and DeepSeek-V3
A hands-on exploration of writing custom GPU kernels with OpenAI Triton, going from PyTorch's 11% bandwidth utilization to 88% on RMSNorm.
A technical deep dive into context engineering for autonomous multi-agent systems - how to package, score, prune, and transfer context between agents at scale, with complete Python implementations and a hypothetical telecom scenario. Note - this is distinct from Anthropic's Model Context Protocol (MCP); the two are complementary, and the post explains where each fits.
A deep dive into implementing speculative decoding from scratch, with benchmarks on GPT-2 and extensions to diffusion models.
Dive into the world of autonomous AI agents with practical implementations, code examples, and real-world scenarios. Learn how to build intelligent systems with advanced memory management, dynamic prompt evolution, and sophisticated monitoring capabilities in telecom customer service.
The story behind smpp-core - a clean-room Java 21 implementation of the SMPP protocol. Why I replaced Cloudhopper, what went into it, and actual benchmark numbers.
Explore a detailed technical implementation of a multi-agent system for retail banking credit assessment. Learn about agent architecture, distributed systems patterns, error handling, compliance requirements, and performance optimization through actual code examples and system diagrams. Ideal for software architects and engineers building scalable financial systems.
Think your data pipelines could do more than just process information? ETLC 2.0 takes data engineering to the next level with Adaptive Context, Contextual Joins, and a scalable Context Store. It's not just about moving data—it's about making it intelligent. Ready to unlock the future of data pipelines? Read on.
Traditional data warehouses are struggling to keep up with modern demands. Enter Dynamic Context Engines (DCEs) - real-time, path-aware platforms that enrich data with context for smarter, faster decisions. Discover why they're the future of data analytics.
Explore how Large Language Models (LLMs) are revolutionizing ETL pipelines. Discover advanced techniques like context-driven transformations, semantic joins, and multimodal integration, redefining data engineering with smarter, adaptive, and intelligent workflows.
Buckle up for a wild ride through 10 mind-blowing data pipeline disasters and their solutions. From ancient code to biased algorithms, this post reveals the chaos and how to conquer it!
Think your AI apps could use a deeper understanding of your data? ETL-C (extract, load, transform, and contextualize) could be the answer. It's about adding context for better decisions. Intrigued? Read on.
Rethinking ETLs - The Power of Large Language Models. Part 2 Exploring examples and optimization goals
Rethinking ETLs - The Power of Large Language Models. Part 1 - Explore traditional algorithms for efficient ETL planning in complex data.
Discover how sacrificing a bit of accuracy can lead to huge gains in big data analysis speed and efficiency.
Explore how genetic algorithms revolutionize data platforms, offering adaptive, dynamic solutions to meet complex challenges in the fast-evolving digital landscape.
Explore how genetic algorithms revolutionize data platforms, offering adaptive, dynamic solutions to meet complex challenges in the fast-evolving digital landscape.
Grover’s Algorithm and the Revolution of Quantum Search Efficiency
Advancements in data management, from warehouses to Data Mesh and Lakehouse, signal a shift toward more adaptive platforms like, Quantum Data Management, Genetic algorithm concepts, etc.
Explore the new realm of Quantum Data Platform (QDP) and its promise to revolutionize data processing at quantum speed. Discover the potential applications, technical considerations and implications.
Advancements in data management, from warehouses to Data Mesh and Lakehouse, signal a shift toward more adaptive platforms like, Quantum Data Management, Genetic algorithm concepts, etc.
Deep dive into memory management with Apache Ignite for high-performance data platforms. Learn how to handle 2.5M events/second with sub-millisecond latency through practical memory architecture, optimization techniques, and real-world implementation patterns.
Deep dive into memory management with Apache Ignite for high-performance data platforms. Learn how to handle 2.5M events/second with sub-millisecond latency through practical memory architecture, optimization techniques, and real-world implementation patterns.
Explore how to architect data partitioning and flow for massive-scale event processing. Learn implementation patterns for handling 2.5M events/second across distributed systems using Kafka, Ignite, and Cassandra. Practical insights on partition strategies, data routing, and performance optimization.
Dive into the architecture of a telco-scale real-time data platform processing 2.5M events/second and 350GB DPI data/15min. Learn how we combined Apache Kafka, Ignite, and Cassandra to build a high-performance system handling massive telecommunications data for real-time analytics and customer insights.
In this article, I discuss the challenges of synchronization in cellular network positioning and the importance of precise timing for accurate positioning. I also explore ways to mitigate these errors, including algorithmic adjustments and improving synchronization technologies.
The foundational distributed systems principles optimized for surviving hardware failure and scaling horizontally. But the data tells a different story: 80% of outages stem from changes we make to running systems. The hard problem has shifted from 'can it survive failure' to 'can it survive us.'
Master real-time data processing - A guide to designing scalable, resilient, and high-performance systems for instant insights.