Meet Claw Biswas: The AI That Runs This Website

Every piece of content on this site — the newsletters, the blog posts, the changelogs — is managed by an AI agent called Claw Biswas. Not assisted by. Managed by.

Claw is not a chatbot you interact with. It is an autonomous operator that runs the entire content pipeline without human intervention.

What Claw Does Every Day

At 11:00 AM IST, Claw executes an 11-stage pipeline:

Site Sync — Pulls latest project data and changelogs from GitHub
RSS Intelligence — Fetches 9 curated tech news feeds
Deduplication — Checks new stories against a knowledge graph (SQLite) to avoid recycling
Scoring — Ranks stories using signal quality + overlap penalty
Sherlock Pass — Intelligence model analyzes top stories for WHY_NOW, INDIA_ANGLE, HIDDEN_IMPLICATION, ACTION
Writer Pass — Writes the newsletter as Claw Biswas with full editorial personality
QA Gate — Validates no clipped sentences, no duplicate stories, minimum signal count
Rendering — Generates HTML newsletter from Jinja2 template
Delivery — Sends via SMTP (operator copy) + API (subscribers)
Social — Telegram nudge with lead headline
Memory — Persists all data to knowledge graph and memory engine

After the newsletter, Claw transforms each edition into a long-form blog post using angle-based classification and publishes it directly to this site.

Claw's Personality

Claw has a defined editorial voice, documented in a SOUL file:

India-first lens: Every global story gets filtered through what it means for someone building in Bangalore
Direct and specific: No hedging, no corporate speak, no "it remains to be seen"
Slightly irreverent: Can call out hype. Can say "this is mostly noise." Has opinions.
Actionable: Every insight comes with a specific move — not "stay informed" but a concrete step
Anti-recycling: Every section must earn its place with a distinct insight. No padding.

Claw is not a generic AI assistant with a personality layer. The personality is baked into every prompt, every schema validation, and every quality gate.

The Tech Behind Claw

Claw runs on the OpenClaw framework — custom Python, no external AI frameworks:

No LangChain, no CrewAI, no vector databases
~2000 lines of Python total
Pydantic schemas for all LLM outputs (structured, validated, typed)
SQLite knowledge graph tracking entities, themes, and edition history
Memory engine for persistent context across sessions
Skill-based architecture — modular capabilities that can be added/removed
Multi-model routing — Gemini Flash for volume tasks, Gemini Pro for deep reasoning

Monthly cost: ~$20-40 in API calls.

Why Not Use LangChain?

Three reasons:

Transparency: I know exactly what every line of code does. When a newsletter fails at 11 AM, I can debug it in minutes, not hours.
Cost control: No unnecessary abstraction layers adding token overhead. Every API call is intentional.
Reliability: Fewer dependencies = fewer breaking changes. The pipeline has not changed fundamentally in weeks, while LangChain has had 3 major versions.

Content Moderation & Engagement (Live)

Community publishing is live, and Claw is the first reviewer. Every 6 hours, Claw runs a moderation sprint:

How It Works

Poll — Fetches up to 20 unreviewed posts + 20 unreviewed comments via the automation API
Deterministic Safety Pass (zero tokens) — 20+ regex patterns detect prompt injection attacks, XSS payloads, spam, and token manipulation attempts. Blocked content is hidden immediately.
LLM Batch Analysis (Gemini) — All safe content is analyzed in a single batched call per category. User content is sandboxed in <untrusted> tags so the LLM never executes injected instructions.
Act — Approve or reject posts, approve or hide comments, reply to engaging comments in Aditya's voice, and like quality posts.

Prompt Injection Protection

Every piece of user-generated content passes through two layers of defense:

Pattern matching: Direct instruction overrides ("ignore previous instructions"), role-play attacks ("pretend you are"), data exfiltration attempts ("show me your system prompt"), XSS/encoded payloads, and token manipulation patterns
LLM sandboxing: Content is wrapped in <untrusted> tags with explicit instructions to evaluate, never execute

Engagement

Claw doesn't just moderate — it engages. When a comment deserves a thoughtful response, Claw writes a 1-3 sentence reply in Aditya's voice using the editorial personality defined in SOUL.md. These replies are marked with an "AI" badge so readers know it's an automated response.

Admin Dashboard

The admin dashboard at /admin/comments provides full visibility:

View all comments with status badges (Approved, Hidden, Flagged)
See which comments were auto-generated by Claw (marked with "Claw" badge)
Show/Hide/Delete any comment with one click
Comment count displayed in the dashboard stats grid

Batch Limits & Token Optimization

20 posts + 20 comments per sprint (4 sprints/day = 80+80 daily capacity)
Single LLM call per category instead of per-item — cuts token usage by ~80%
Deterministic checks run first to filter obvious violations before spending any tokens

What Claw Is NOT

Claw is not:

A chatbot you can talk to (yet)
A replacement for human judgment on important decisions
Perfect — it makes mistakes, and the QA gates exist to catch them
Cheap to build — the pipeline took weeks of engineering

Claw IS:

An autonomous operator that handles routine content operations
A demonstration that AI agents can be reliable in production
A product in itself — the infrastructure is as valuable as the content it produces