June 2026 Summaries

PostHog initially designed its data warehouse to delay the need for hiring data engineers in small companies, but as businesses grew, they faced challenges with the scalability and flexibility of their system. The core issue was the transition from using PostHog as the primary analytics tool to needing more complex data engineering solutions, which led to exporting data to external warehouses and utilizing PostHog mainly as middleware. Challenges arose from their use of a multi-tenant ClickHouse cluster, which struggled with general-purpose warehousing needs, and the lack of sophisticated data tooling. To address these issues, PostHog rebuilt its system using DuckDB, creating single-tenant instances that integrate with modern data tools through a Postgres Wire protocol endpoint, and introduced DuckHog for efficient local data handling. This new architecture allows businesses to seamlessly integrate their event data with external data sources, providing a unified and trustworthy data context essential for AI-driven product development and enabling more effective and insightful agentic workflows.

AI observability tools are essential for understanding and monitoring the performance of LLM-powered applications by capturing inputs, outputs, latency, token usage, and more. These tools function like glass walls on a kitchen, providing transparency into every request and response to help evaluate output quality. The text discusses various AI observability tools, including PostHog, Langfuse, LangSmith, Arize Phoenix, Braintrust, Weights & Biases Weave, Opik, LangWatch, and OpenLLMetry, each with specific strengths, pricing models, and use cases. Key features of these tools include tracing, logging, cost tracking, prompt management, evaluations, and datasets for regression testing. The choice of tool depends on factors such as team size, pricing preferences, the need for open-source solutions, and integration with existing observability stacks. Additionally, the text advises on considerations like avoiding per-seat pricing for growing teams, self-hosting for cost predictability, and the relevance of OpenTelemetry-based observability for flexibility and avoiding vendor lock-in.

Langfuse is an open-source platform designed for LLM observability, offering features such as tracing, prompt management, evaluations, and cost tracking to help developers monitor their LLM applications in production. However, not all teams find Langfuse suitable, prompting a comparison of alternatives like PostHog, Braintrust, LangSmith, Arize Phoenix, and Weights & Biases Weave. PostHog is notable for integrating AI observability with product analytics and session replay, making it ideal for teams wanting comprehensive insights into both AI performance and user behavior. Braintrust focuses on evaluation-driven development, offering extensive evaluation and experimentation tools. LangSmith stands out for its deep integration with LangChain and LangGraph, automatically capturing traces with minimal setup. Arize Phoenix emphasizes OpenTelemetry-native instrumentation and agent tracing, while Weights & Biases Weave extends its existing model training and tracking workflows into LLM observability. Each alternative has unique strengths, with PostHog recommended for those seeking a blend of AI observability and product development tools, and Langfuse remaining a strong choice for focused prompt management and detailed agent tracing.

The author undertakes the ambitious task of rewriting PostHog's SQL parser using AI assistance through multiple long-running Claude Code sessions, resulting in a significant performance boost and a new parser that is 454 times faster than its predecessor. The undertaking involves creating a "hand"-written recursive-descent parser with a Pratt expression core, achieving complete agreement with the existing C++ ANTLR-based parser for realistic SQL queries. The approach combines property-based testing with code-generation tools to create inputs, alongside prompt engineering to address AI's brittle fixes, ultimately leading to a robust parser that can withstand rigorous testing. The new parser was developed in Rust, leveraging AI capabilities to achieve a task that would have traditionally taken months, illustrating a potential shift in how parsers might be developed in the future, with AI-generated high-performance parsers complementing traditional parser generators like ANTLR.

A scout in PostHog is a small, autonomous agent designed to analyze data within the platform to identify patterns and emit actionable signals, forming a crucial component of PostHog's self-driving functionality. Operating as a skill primarily written in markdown, scouts translate complex business logic into easily readable English, allowing both human users and agents to understand and act upon the insights generated. Scouts work by running scheduled analyses on data slices, emitting signals when noteworthy patterns are detected, which are then processed into reports and potentially trigger automated fixes or improvements via PostHog's self-driving pipeline. The scouts' effectiveness lies in their ability to discern meaningful findings, avoiding false positives by maintaining a memory of past data interactions, and thus only surfacing issues when they meet specific thresholds of significance. This system, which supports flexibility and iterative improvement, enables users to focus on strategic decision-making by automating routine tasks and improving data pipelines’ efficiency and reliability.

PostHog is developing an automated product improvement pipeline that ingests vast amounts of product data, groups signals into actionable problems, and uses AI agents to propose fixes, resulting in automated pull requests (PRs) ready for review in GitHub. The system addresses the inefficiencies of traditional product improvement loops by employing a classifier to filter out malicious signals and aligning signals with actual problems using a language model (LLM). Once signals are grouped, a research agent analyzes the problems, checks their actionability, and, if actionable, writes and iterates on code solutions. This approach is designed to transform the traditional workflow by reducing manual interventions and enabling automated, data-driven code improvements, ultimately aiming for a self-driving product development process that continuously learns from each outcome.

AI observability tools are essential for detecting issues like latency, token costs, and model failures in AI systems before they impact users. These tools offer various free tiers, which differ significantly in terms of event limits, data retention, and user seats, making it crucial to choose the right one based on specific needs. PostHog stands out with its generous free tier of 100,000 events per month and unlimited seats, integrating AI observability with product analytics and error tracking. Langfuse offers deep LLM-native workflows with prompt management and evaluation capabilities, while Traceloop provides OpenTelemetry-native ingestion for vendor-neutral setups. Other platforms like Arize, Lunary, HoneyHive, LangSmith, and Braintrust cater to various team sizes and needs, emphasizing features like OpenTelemetry integration, evaluation depth, and self-hosting options. The right tool depends largely on the specific requirements of the team, such as project scale, need for open-source solutions, or integration with existing platforms like LangChain or Datadog.

In the current digital landscape, "build mode" represents a transformative approach where individuals from various professional backgrounds, such as product engineers, managers, marketers, and support staff, collaborate and contribute to product development without being confined to traditional roles. This mode is facilitated by tools like PostHog's AI, which allows users to interact with data and execute tasks typically reserved for engineers, such as opening pull requests, analyzing user data, and creating automated insights. The concept emphasizes empowerment and capability enhancement across roles, enabling everyone to participate in the creation, marketing, and improvement of products. By leveraging AI, non-engineers can perform actions like coding, analyzing data, and making informed content decisions, demonstrating that the integration of AI into workflows allows for a more dynamic, cross-functional approach to product development and problem-solving. This shift not only democratizes access to technical processes but also fosters an environment where ideas can be swiftly transformed into reality, thus accelerating innovation and collaboration.

Social media metrics such as impressions, reach, views, engagements, followers, and engagement rates are often misleading indicators of success, as they can be easily manipulated and lack context about how audiences genuinely engage with content. Instead of focusing on these quantitative metrics, brands should prioritize consistent brand representation and creating content that resonates with their ideal customer profile (ICP). By doing so, companies can foster genuine engagement and cultivate a loyal audience. It is crucial for social media teams to enable content creation across the organization and empower employees to share their own content, as this approach encourages authentic interaction and reinforces brand identity. Engaging with similar brands and understanding industry-specific contexts can also guide effective social media strategies, ensuring that content remains relevant and appealing to the target audience.

A marketer, who was once unable to code, describes their journey in creating "Joe-OS," a desk dashboard software for a Pimoroni Presto device using MicroPython, which integrates an environment sensor and various interactive panels. The project, aided by AI tool Claude, includes features such as news readers, a daily joke, and a medieval-themed text adventure game called Vigil, which involves decision-making with moral choices. Despite not being an engineer, the creator emphasizes how AI has bridged the gap between having an idea and implementing it, while analytics via PostHog have enhanced their understanding of user interactions and device performance. This newfound insight has allowed them to make data-driven decisions about feature updates and user interface tweaks, illustrating the evolving role of AI and analytics in non-technical projects.

At PostHog, the roles of customer success managers (CSMs) and technical account managers (TAMs) are redefined to focus on genuine customer engagement and proactive problem-solving rather than traditional, bureaucratic practices typical in many SaaS companies. Their days involve a mix of technical troubleshooting, like resolving data mapping bugs and optimizing customer implementations, and creative customer interactions, such as delivering donuts and crafting personalized emojis. PostHog values authenticity and technical expertise, encouraging its team to directly address customer issues, often leading to increased trust and deeper product adoption. By prioritizing customer success through unconventional yet effective methods, PostHog fosters a vibrant, high-agency work culture that aligns genuine engagement with tangible outcomes.

AI observability addresses the unique challenges of monitoring AI applications, which traditional monitoring tools cannot adequately handle due to AI's variability in outputs and behaviors. It encompasses tracking AI features such as prompts, responses, costs, latency, errors, and output quality, filling a gap that traditional Application Performance Monitoring (APM) tools leave since they are designed for consistent software operations. AI observability uses unique data models centered on generations, traces, and spans to provide insights that traditional tools can't, such as semantic clustering of model outputs. This approach is crucial for anyone deploying AI features to assess performance, debug errors, and optimize costs effectively. Various tools cater to the AI observability landscape, like PostHog, Langfuse, LangSmith, Datadog, Portkey, and Arize Phoenix, each offering distinct features such as tracing, cost tracking, error capture, and evaluation frameworks. These tools differ in integration capabilities, pricing models, and data ownership options, with some offering open-source solutions for self-hosting. Ultimately, choosing the right tool depends on factors like existing infrastructure, desired level of integration, and specific observability needs.

DuckDB and SQLite are both lightweight, embedded databases that share similar design principles, but they cater to different use cases, with SQLite being a transactional OLTP database and DuckDB being an analytical OLAP database. SQLite is known for its simplicity, minimal configuration, and suitability for transactional local storage, making it ideal for applications like mobile apps and browsers. It operates as a single-file, in-process database without the need for a client-server infrastructure, although it has limitations in terms of concurrency and horizontal scalability. On the other hand, DuckDB excels in analytical queries by utilizing columnar storage and vectorized execution, allowing it to efficiently handle large datasets and complex queries without extensive setup. While it shares SQLite's lightweight, serverless nature, DuckDB is designed for analytical workloads and can work with massive data files, making it portable and powerful for analytical tasks. Despite their differences, both databases embody the philosophy of being simple, in-process libraries that prioritize ease of integration over distributed architecture, with SQLite focusing on transactional capabilities and DuckDB on analytical performance.

The text examines the contrasting characteristics of Postgres and DuckDB, highlighting their distinct design philosophies and applications. Postgres is a general-purpose, client-server, row-based OLTP database suited for handling transactions and maintaining a central system database, while DuckDB is an embedded, column-based OLAP database optimized for analytical processing. The differences extend to their memory structures, query execution models, and scalability methods, with Postgres using heaps and B-trees for efficient data access, and DuckDB employing columnar storage and zone maps for large-scale data analysis. DuckDB's decoupled storage allows it to process vast datasets efficiently without requiring extensive local resources, whereas Postgres provides robust multi-user access and persistent data management. The text also describes how PostHog, a single company, employs both databases for different purposes: Postgres for application state and metadata, ClickHouse for scalable analytical data, and DuckDB for querying data warehouses, enhanced by their Duckgres project to integrate DuckDB with Postgres-compatible tools.

PostHog has introduced a beta version of their Slack app, designed to streamline the process of addressing minor software issues by allowing users to directly tag @PostHog with requests like "fix this" or "build that." This integration uses the context of product data and adheres to repository rules to create draft pull requests, conduct checks, and respond to review comments, making it accessible to non-engineers across various roles such as sales, marketing, and customer support. The app has demonstrated its utility through diverse applications, from generating code for new features to preparing for user interviews and updating company documentation. It maintains a rigorous review process, ensuring that all changes go through a thorough validation before merging, and supports both analytics and coding tasks by leveraging PostHog AI credits for its operations. The app stands out by connecting directly with product data, enabling it to diagnose issues from data analytics and propose solutions without requiring users to switch between different tools.

AI observability, particularly for small teams and solo developers, focuses on monitoring the internal workings of AI features within an application, such as prompts, responses, and associated costs, without the need for over-complicated setups typically designed for large-scale enterprises. This guide emphasizes the importance of establishing basic AI observability from the start to avoid unexpected costs and improve debugging processes, suggesting that minimal initial setup should include tracing LLM calls and tracking costs to prevent financial surprises and ensure quality outputs. Unlike traditional APM tools, which may confirm an API call's success without evaluating its content or cost, AI observability tools provide insights into the quality and economics of language model operations. As a product matures, additional layers such as automated evaluations and user feedback can be integrated to enhance the observability framework, allowing for a more comprehensive understanding of AI performance and its impact on user retention. The guide suggests leveraging existing tools like PostHog, which offers a free tier, to efficiently handle tracing, cost tracking, and error capture without needing custom-built infrastructure, thereby enabling teams to focus on product development rather than maintenance of observability systems.

In September 2025, OpenAI acquired Statsig and appointed its founder Vijaye Raji as CTO of Applications, marking a significant shift in the experimentation and feature management landscape. By May 2026, Amplitude had announced a strategic partnership to take over Statsig's brand, platform, and customers, while the original team remained at OpenAI, prompting teams relying on Statsig to explore alternatives based on specific needs like rapid releases, analytics depth, and no-code experiments. Notable alternatives include PostHog, which offers a comprehensive developer platform combining feature flags, experimentation, and product analytics, making it a strong contender for startups and engineering-led teams. LaunchDarkly stands out for enterprise governance and automation, while Amplitude provides deep product analytics suited for non-technical users. For marketing teams, Optimizely and VWO offer robust no-code experiment capabilities, with VWO focusing on conversion optimization through heatmaps and personalization. GrowthBook appeals to data scientists with its open-source, warehouse-native approach, and Kameleoon offers developer-focused optimization with AI assistance. Each alternative presents unique strengths, catering to diverse organizational needs in the evolving landscape of feature management and experimentation tools.

During a team offsite in Lisbon, an AI agent was utilized to analyze slow queries in ClickHouse, revealing a three-year-old bug related to improper use of primary keys with timestamp filters. This bug led to inefficient query performance, but the AI-driven analysis identified a fix that reduced granule scans by 62% and improved query speed significantly. The approach was inspired by Andrej Karpathy's concept of "autoresearch," which involves iteratively testing changes to improve system performance without the biases inherent in human coding practices. The team structured their analysis into campaigns with specific optimization goals and utilized a small coding agent, pi, to automate the process. This led to the discovery that the toTimeZone() function was preventing effective partition pruning. By modifying the query to allow the ClickHouse planner to see a bare timestamp, significant performance improvements were achieved. Going forward, the team plans to automate this process further by fetching slow queries from logs and running them through a pipeline that uses PostHog Code to implement and test improvements, aiming to make such optimizations a routine, automated task.