LLM Apps on Narasimha Karthik J

whoblinked

Wed, 01 Jul 2026 00:00:00 +0000

What it is #

whoblinked watches your competitors’ catalogues and tells you what they launched, discontinued and repriced. I co-founded it and have been CTO since July 2026. The code is private.

I built it end to end, from the data model and API contracts to the React product. It is multi-tenant.

How it works #

Extraction with fallbacks. Each scan tries the store’s API first, then structured markup on the page, then a headless browser.
Block detection. A scan that was blocked or failed is recognised as such, so it never shows up as a change in a competitor’s catalogue.
LLM features. Competitor discovery, gap analysis and ownership research. Model output is checked against live search results before a user sees it.
Cost and release. Every LLM call has its cost tracked, and an automated test-and-release pipeline gates deploys.

LockIn: Browser Focus Tracker

Thu, 02 Apr 2026 00:00:00 +0000

Code on GitHub

What it is #

You start a session with a goal, such as “fix auth bug in login flow”. The extension records each page you open. An LLM labels each visit on-task, distraction or ambiguous relative to that goal. If you drift, an overlay nudges you and offers to take you back. When the session ends you get a focus report.

How it works #

Extension (Manifest V3). Sends each page load, with its URL, title and page text, to the backend, and polls a nudge endpoint every 60 seconds.
Backend (FastAPI). Stores the visit and returns at once. Classification runs as a background task, so browsing never waits on the model.
Classification. A RocketRide pipeline sends GPT-4o the goal, URL, title and the first 2,000 characters of the page, and asks for a JSON label and reason.
Graph storage (Neo4j). Sessions, tasks, visits, sites and nudges are nodes. Visits link in order with NEXT edges, and sites link to tasks as ON_TASK_FOR or DISTRACTION_FROM.
Report and Q&A. The report shows a focus score, a color-coded timeline, a per-site breakdown and LLM insights. A second pipeline answers plain-language questions about past sessions by querying the graph.

Neo4j runs in Docker Compose, and a seed script loads a demo session.

Zettelkasten Automation

Fri, 16 Jan 2026 00:00:00 +0000

Code on GitHub

What it is #

A command-line tool for an Obsidian vault. It reads long source notes (decoded papers, articles), pulls out the reusable ideas as atomic notes, links them, and keeps a local semantic index. It can then gather the notes around a source or a note into one bundle to hand to a model or read in one place.

It never deletes a source. A processed source gets a Status: #processed marker.

How it works #

zettel process handles each unprocessed source:

Skips notes that are already processed or too short.
Splits long sources by heading, then by paragraph.
Runs structured extraction per chunk and validates the output.
Drops candidate concepts that duplicate an existing note, by normalized title or by embedding distance.
Writes new notes from the vault template with Tags, See Also (semantic neighbours) and References (the source).
Indexes the notes in ChromaDB.

Each source ends with a named outcome, such as processed, too_short, provider_error or invalid_output.

zettel bundle collects the notes from a source or around one note, following tag links to a set depth. A dry run refuses to use a stale index rather than return wrong neighbours.

Details #

LLM providers: OpenAI, Anthropic, Gemini and Ollama. Embeddings: OpenAI or a local sentence-transformer. The index is scoped to the embedding provider and model, so switching models does not mix vectors.
--dry-run on every command that writes.
The research-vault skills in claude-kit use it for semantic search over zettels.