Redar: AI-Powered Summaries for KAP Disclosures and Open Sources

Financial information in Turkey moves fast, and much of it arrives in dense, inconsistent formats. Redar is a private decision-support and monitoring panel for following Kamuyu Aydınlatma Platformu (KAP) disclosures and selected open sources. It helps turn the updates we choose to follow into compact summaries and notifications in the places we use them.

A useful system does not have to become a public product

Redar started with a practical need shared by a friend. The question was not how to build another finance dashboard. It was how to follow the disclosures and sources that matter to us without repeatedly reading irregular source material from scratch.

It is still running, but it is deliberately small: a private panel used by my friend and me. The data rights behind the project do not allow public data publication, so opening it as a public service is not its intended path. That boundary is useful to state plainly. Software can solve a real problem for a defined group without needing to become a SaaS product or a launch story.

The panel tracks KAP disclosures and selected open sources, supports search and filtering, and sends Telegram notifications in a format that fits this narrow workflow. The aim is not to make every disclosure look equally important. It is to reduce the effort required to review the ones we have chosen to monitor.

From irregular source material to a reviewable signal

KAP disclosures do not arrive in one dependable structure, and the HTML markup can be difficult to work with. Redar therefore uses two paths:

• Deterministic parsing for any source that arrives in a standard structure.
• LLM-assisted extraction and summarization when the structure breaks down.

The deterministic path remains preferable where it fits: it makes the extraction rule visible and repeatable. The LLM path is an assistive fallback for messy documents, not a claim that every source can be interpreted perfectly. The result is a compact web view and, for selected updates, a Telegram notification that still leaves room for the original disclosure when more context is needed.

Telegram is a first-class delivery channel because the decision is often whether to look closer now, not whether to spend time inside another dashboard. That is a modest use case, but a useful one.

Tech stack at a glance

• Backend: Fastify
• Frontend: Nuxt v4
• Data: MariaDB
• Caching: Redis
• LLM: OpenAI API (different models for different tasks—e.g., Nano, Mini—no custom training yet)
• Networking/Infra: Traefik reverse proxy, Docker Swarm on my existing VPS
• CI/CD: GitHub Actions
• Search & Ops: Server-side filtering and querying with latency kept tight through Redis and careful parsing paths

This stack supports ingestion, normalization, search, and notification for a small private workflow.

The constraint is part of the design

Redar is private by design. The underlying data licenses do not allow public data publication, and the intended users are already clear. That makes a focused internal-style panel more honest than presenting it as a public financial product waiting to launch.

The project also reinforces a practical implementation lesson. Deterministic parsing and LLM assistance are not competing ideologies. The first is valuable when the input is dependable. The second can help when it is not. In both cases, the system should make review easier rather than pretend to replace judgment.

For now, Redar remains useful in that limited form. Keeping the scope aligned with its data rights and its real users is a product decision, not an unfinished launch plan.