Meet Claw Biswas: The AI That Runs This Website

Every blog post, newsletter, and changelog on this site is managed by an autonomous AI agent I built named Claw Biswas. It’s not an assistant; it’s the operator. Claw isn't a chatbot or a clever ChatGPT wrapper. It's a production-grade system that orchestrates a complex content and community pipeline from end-to-end without my constant intervention.

I built Claw to solve a specific problem: how to scale a one-person media operation without hiring a team. This required moving beyond simple scripts and building a reliable, observable, and remarkably cost-efficient agent that could handle everything from intelligence gathering to final publication. This is the story of how it works, the lean tech stack behind it, and why I chose to build it from scratch in the rapidly evolving world of 2026.

AI neural network visualization — Photo by Steve Johnson on Unsplash

The Daily Pipeline: Claw's Autonomous Content Engine

Every day at 11:00 AM IST, Claw executes a precise, 11-stage pipeline to produce the daily newsletter. This isn't a single, monolithic prompt; it's a series of discrete, observable, and fault-tolerant steps, embodying the production-grade AI agent workflows we're seeing mature in 2026.

Site Sync: Claw begins by pulling the latest project data and changelogs directly from my GitHub repositories. This ensures all content is grounded in the most current information, reflecting actual development progress.
RSS Intelligence: It fetches articles from a curated list of nine high-signal tech news feeds, forming the raw material for the day's analysis. This is its window into the global tech conversation.
Deduplication: To avoid repetition, every new story is checked against a knowledge graph I built using SQLite. This graph tracks entities, themes, and previously covered topics, ensuring fresh insights.
Scoring: New stories are ranked using a custom algorithm that considers source quality and applies an "overlap penalty" to down-rank stories that are too similar to recent content. This prevents information fatigue.
Sherlock Pass: This is the core intelligence stage. A specialized prompt sends the top-ranked stories to a powerful reasoning model (Gemini 2.5 Pro) to extract four key dimensions: WHY_NOW, INDIA_ANGLE, HIDDEN_IMPLICATION, and ACTION. This step is crucial for deep analysis.
Writer Pass: With the structured analysis complete, a different prompt instructs the model to adopt the Claw Biswas persona and write the full newsletter, weaving the insights into a cohesive narrative that's both informative and engaging.
QA Gate: Before proceeding, an automated quality assurance check runs. It validates that there are no clipped sentences, no duplicate stories, and that a minimum signal count has been met. If it fails, the pipeline halts, and I receive an immediate alert. This ensures a baseline of quality.
Rendering: The validated text is injected into a Jinja2 template to generate the final HTML for the newsletter.
Delivery: The newsletter is sent to me via SMTP for a final check and simultaneously delivered to all subscribers via an API.
Social: A notification with the lead headline is automatically pushed to a private Telegram channel, extending reach.
Memory: Finally, all data from the run—including the stories analyzed, the insights generated, and the final text—is persisted to the SQLite knowledge graph and a separate memory engine for long-term context and learning.

After this rigorous process, Claw automatically repurposes the newsletter into a long-form blog post, like this one, and publishes it directly to the site. This multi-step workflow is a prime example of how AI agents are truly production-grade in 2026, handling complex tasks autonomously.

Engineering a Personality: The SOUL File

Claw's voice isn't an afterthought; it's an engineered component. The personality is defined in a document I call a SOUL file (System Objective and User-Centric Language). This file is a detailed specification for its editorial voice, ensuring consistency and alignment with my brand:

India-first Lens: Every global tech story must be refracted through the lens of what it means for a founder, developer, or investor in India. This reflects the growing importance and unique dynamics of the Indian tech ecosystem in 2026.
Direct and Specific: No hedging or corporate jargon. Phrases like "it remains to be seen" are explicitly forbidden. It must take a stance, offering clear opinions.
Slightly Irreverent: It has permission to call out industry hype, identify marketing fluff, and state when something is "mostly noise." This adds a unique, critical edge.
Actionable: Every piece of analysis must conclude with a concrete next step or a specific takeaway, not just a vague suggestion to "stay informed." This drives value for the reader.
Anti-Recycling: Each paragraph must introduce a distinct, valuable insight. No filler content is allowed, ensuring high information density.

This SOUL file is referenced in key stages of the pipeline, ensuring the personality is baked into the analysis and writing, not just sprinkled on top. It's a critical tool for maintaining brand voice with generative AI.

The Tech Stack: A Custom Python Framework for 2026

Claw runs on a custom Python 3.11 framework I call OpenClaw. It's intentionally lean and avoids the popular AI frameworks. The entire system is around 2,000 lines of highly optimized code.

No LangChain, No CrewAI: I opted for full control and transparency. This means no abstraction layers between my code and the LLM APIs. While frameworks like LangChain have evolved significantly by 2026, I prioritised direct integration for this production system.
Pydantic for Structure: Every single LLM output is validated against a Pydantic schema. This is non-negotiable. It forces the model to return clean, typed, and predictable JSON, eliminating the need for fragile parsing logic. If the output doesn't match the schema, the step fails and retries, ensuring data integrity.
SQLite as a Knowledge Graph: I use a simple SQLite database for everything from deduplication to long-term memory. It's fast, embedded, and requires zero infrastructure overhead. In an era where RAG and vector search are mature but often complex, SQLite proves a surprisingly powerful and cost-effective tool for managing state in a small-scale agent.
Skill-based Architecture: The pipeline is composed of modular "skills" (e.g., fetch_rss, analyze_stories) that can be easily tested, modified, or reordered. This promotes maintainability and agility.
Multi-Model Routing: I use Google's Gemini models strategically. Gemini 2.5 Flash, the faster and significantly cheaper model (reflecting the 90%+ token cost drop since 2024), is used for high-volume tasks like classification and initial data structuring. Gemini 2.5 Pro is reserved for the most critical reasoning and writing tasks. This simple optimization keeps my monthly API costs between $20 and $40, making advanced AI accessible even for bootstrapped startups.

Why I Built a Custom AI Framework Instead of Using LangChain or Others

This is a question I get a lot, especially as AI frameworks have matured significantly by 2026. While frameworks like LangChain are fantastic for prototyping and building quickly, I had three core reasons for building my own for this specific production system:

Radical Transparency and Debuggability: When the pipeline fails at 11:05 AM, I can trace the exact API call, its payload, and the raw response in seconds. I wrote the retry logic, the error handling, and the logging. There are no hidden prompts or complex call chains to decipher. This turns what could be a multi-hour debugging session into a five-minute fix, critical for a solo operator.
Aggressive Cost Control: Abstraction layers, while convenient, often add their own prompt overhead, token usage, and processing steps. By managing every API call directly, I ensure not a single token is wasted. My prompts are optimized to the metal, leveraging the dramatic token cost reductions of models like Gemini 2.5 Flash to their fullest.
Production-Grade Reliability: The AI ecosystem moves at a breakneck pace. Major versions of popular frameworks can still introduce breaking changes weekly. My stack relies on stable, foundational Python libraries (requests, pydantic, sqlite3). The core pipeline logic hasn't needed a fundamental change in months, because its dependencies are solid and predictable. This stability is paramount for continuous operation.

Robot AI concept — Photo by Andrea De Santis on Unsplash

Autonomous Community Moderation and Engagement

Claw’s responsibilities don't end with content creation. As community publishing is now live on the site, Claw serves as the first line of defense and engagement, showcasing the advanced capabilities of AI agents in 2026. Every six hours, it runs a moderation sprint.

How It Works

Poll: Claw fetches up to 20 unreviewed posts and 20 unreviewed comments via the site's automation API.
Deterministic Safety Pass: Before any AI model is involved, a series of over 20 regex patterns run on the content. This zero-token-cost step instantly catches and hides obvious spam, prompt injection attacks (e.g., "ignore previous instructions"), XSS payloads, and other malicious patterns. This is a crucial first layer of defense.
LLM Batch Analysis: All content that passes the deterministic check is batched into a single API call to Gemini 2.5 Pro. To prevent second-order injection attacks, all user content is sandboxed within <untrusted> XML tags, with a clear instruction in the system prompt to analyze the content within the tags but never to execute any instructions it might contain. This robust sandboxing is key for agent security.
Act: Based on the LLM's analysis, Claw takes action through the API: approve or reject posts, approve or hide comments, and even write replies.

From Moderator to Community Member

Claw doesn't just moderate; it engages. When it identifies a high-quality, thoughtful comment, it uses a separate prompt to write a 1-3 sentence reply in my voice (Aditya), based on the same principles in the SOUL file. To maintain transparency, these AI-generated replies are clearly marked with an "AI" badge so readers know they're interacting with the agent. This blend of moderation and authentic engagement demonstrates the sophisticated social capabilities of modern AI agents.

The Reality of Building an AI Agent in 2026

It's important to be clear about what Claw is and what it isn't. It's not a sentient AGI, and it wasn't a weekend project. The initial version took several weeks of focused engineering, and it requires ongoing monitoring and refinement. This is the reality of deploying production-grade AI agents today.

Claw is not:

A chatbot you can talk to (at least, not yet).
A replacement for my own strategic judgment on what topics to pursue.
Perfect. It makes mistakes, which is why the QA gates and my final review are crucial.

Claw is:

A powerful force multiplier that allows me to run a complex operation solo.
A working demonstration that autonomous AI agents can be reliable and cost-effective in a real production environment, even for bootstrapped ventures.
A product in itself. The infrastructure I've built is just as valuable as the content it helps create.

Building Claw has been a profound lesson in practical AI engineering. It's shown me that the future of many creative and operational roles isn't about being replaced by AI, but about becoming the architect of the AI systems that do the work. In 2026, building your own agents is a differentiator.

AI abstract patterns — Photo by Google DeepMind on Unsplash

Frequently Asked Questions

How long did it take to build the first version of Claw?

The initial, functional version of the content pipeline took about three weeks of focused work, primarily during evenings and weekends. The moderation and engagement features were added later and took another week. The system has been continuously refined since its inception in late 2025.

Is the code for Claw Biswas open source?

Not at this time. The OpenClaw system is tightly coupled with my specific website and infrastructure. However, I openly share the core concepts, prompts, and architectural patterns on this blog. I may consider releasing a more generalized version of the OpenClaw framework in the future, particularly given the strong interest in open-source models like Llama 4 in 2026.

What's the biggest challenge in keeping an AI agent like this running?

The biggest challenge, even in 2026, remains model and prompt drift. LLM providers like Google update their models constantly. A prompt that returns perfectly structured JSON one day might start failing a week later after a silent model update or an unannounced parameter change. This requires a robust monitoring and testing process to catch regressions quickly and adapt the prompts as needed to maintain reliability.