Skip to content
ByteBrief
In development

Understand what changed. Trace it to the source.

ByteBrief is building an intelligence layer that connects news, research, and podcasts, tracks claims over time, and turns fragmented information into briefings you can verify.

Explore Intelligence LayerBrowse the news

Intelligence Layer

In development

News articles, research papers, podcasts, and public statements often describe the same development in different ways. Intelligence Layer is designed to connect them in a persistent, source-linked knowledge base, so each briefing can show what changed and where every claim came from.

  • Connect information across sources

    Articles, papers, podcast transcripts, and public statements about the same development will be linked, so related coverage reads as one thread.

  • Track claims and how they evolve

    Each claim will be stored with its source and date, then marked as later evidence supports, revises, or contradicts it.

  • Understand what changed, with supporting evidence

    Briefings will start with what is new since your last one, and every change will point to the sources behind it.

Illustrative example — not a live output.

Briefing

Illustrative example: Compact AI models

Since your last briefing, 6 days ago

3 changes from 4 sources

  • Revised claim

    Previously: A lab's new compact model matches much larger models on a standard math benchmark. (Source 1)

    Now: An independent replication finds the gap closes only when the compact model gets extra reasoning time. (Source 2)

  • Contradiction

    The lab's announcement says no synthetic data was used in training. (Source 1) A later preprint reports synthetic math problems in the training mix. (Source 3)

    Both claims stay visible until new evidence settles the question.

  • New development

    In a podcast interview, one of the preprint's authors says the full evaluation code will be released. (Source 4)

    Linked to the benchmark claim

Sources

  1. [1]Lab announcementPublic communication, 9 days ago
  2. [2]Report on an independent replicationNews article, 4 days ago
  3. [3]Research preprintResearch paper, 3 days ago
  4. [4]Interview with a preprint authorPodcast, 2 days ago, at 31:40

Harness Lab

In development

Better agents, measured.

Harness Lab is our experimental environment for testing how instructions, tools, memory, and workflows affect agent performance.

Each variation runs on the same repeatable tasks, so results can be compared directly. The agent designs that hold up are the ones we plan to use in Intelligence Layer.

What we vary

  • Instructions
  • Tools
  • Memory
  • Execution workflows

What we compare

Verified task success
Whether the task was actually completed, checked independently of the agent's own report.
Reliability
How consistently the same setup succeeds across repeated runs.
Latency
How long each task takes to reach a verified result.
API cost
What each run costs in model and tool calls.

Building with Claude

We plan to use the Claude API for structured extraction, reasoning across sources, and research agents. Harness Lab will evaluate how each of them performs.

Structured extraction
Turning articles, papers, and transcripts into consistent records of claims, entities, dates, and sources.
Reasoning across sources
Comparing what different sources say to find where they agree, disagree, or change.
Research agents
Gathering and checking evidence for open questions, then reporting what was found and where.

Built so far

Intelligence Layer builds on a news pipeline we built and deployed in 2025.

  1. May 2025

    News ingestion and AI summaries

    Feed fetchers for MIT Technology Review, Hacker News, and Futurism store articles in MongoDB, and a summarization step adds a summary and bullet points to each one.

  2. May 2025

    Daily digests by email and Telegram

    A scheduled job collects the day's articles and sends them as a digest by email and to Telegram.

  3. May 2025

    Web app with accounts and search

    A Next.js reader on a FastAPI backend, with sign-in, topic preferences, topic filters, and full-text search.

  4. June 2025

    English and Turkish

    A bilingual interface, Turkish translations of summaries, and a Turkish edition of the digest.

  5. Now

    Intelligence Layer and Harness Lab

    In development, starting from this pipeline.

The 2025 news reader is temporarily offline while it moves to a new backend.

Founder

Founded by Bilgin Koçak, an AI/ML engineer with experience building production financial AI and information-extraction pipelines.