Farasu

Client

Own product

Channel

t.me/cognitivedgebyp

Role

Solo — pipeline and editorial rules

Year

2026

Farasu (فراسو) is an AI-run Telegram channel that reads the day's psychology, business and technology reporting, scores it, and publishes only what clears the bar — in Farsi, credited, up to five times a day, with no one at the keyboard.

Join the channel on Telegram →

The Problem

Staying current across three fast-moving fields — psychology, business/MBA, and tech — means scanning dozens of sources every day, then summarizing, translating, and crediting each one. Done by hand it's hours of work, easy to fall behind on, and impossible to keep consistent.

Automating it naively makes a different problem. A pipeline that publishes the freshest item on a schedule will publish something every single time, including on days when nothing worth reading came out. A channel that posts filler to hit a quota trains its audience to scroll past it.

The Solution

Farasu is an end-to-end pipeline that runs on a schedule and publishes to a Telegram channel. Each run:

• Gathers fresh articles from that topic's feeds — TechCrunch, The Verge and Hacker News for tech; PsyPost and ScienceDaily for psychology; MIT Sloan Management Review and Knowledge at Wharton for business. Every feed is set to degrade rather than fail, so one dead source doesn't take the run down with it.
• Shortlists the freshest fifteen, drops anything already published, and packs them into a single request — so scoring an entire shortlist costs one model call, not fifteen.
• Scores each candidate on substance, relevance, credibility, novelty and actionability, weighted 30/25/20/15/10.
• Decides. The best candidate publishes only if it scores 60 or higher. Nothing clears the bar, nothing goes out.
• Fetches the full article, not just the RSS excerpt, and keeps whichever body text is longer — feeds vary wildly in how much they include.
• Writes the post in Farsi with Google Gemini Flash: a hook, a short summary, a "why this matters" line, and a discussion question.
• Translates the whole article into Farsi and publishes it to Telegra.ph, so the channel post carries a link that opens as Telegram Instant View — the full piece, readable in-app, without leaving Telegram.
• Publishes to the channel, then records the outcome — the posted link is written to the dedup history after the message ships, never before.

n8n orchestrates all of it on the schedule below. A separate error-monitor workflow is wired into every content workflow as its error handler, so a failed run sends a direct message instead of failing silently.

The publishing schedule

Design Decisions

Telegram over a standalone app or site. The audience already has Telegram open daily — publishing there instead of building a destination app removes the single biggest adoption barrier: asking someone to install and check yet another thing.

An editorial gate, not just a sort. The earlier version took the freshest article and posted it. That guarantees a post, which sounds like a feature until you read a week of them — quiet days produce filler, and filler is what teaches people to stop opening a channel. Scoring a shortlist and publishing nothing below the threshold means the channel's silence carries information: if it posted, something was worth posting.

Scoring the whole shortlist in one call. Fifteen candidates could be fifteen model calls per slot. Packing them into one request makes the gate affordable enough to run on every branch, every day — the cost structure is what decides whether a quality filter survives contact with a real schedule.

Crediting as a hard rule, not a toggle. Every summary keeps a visible link back to its original source, enforced in the pipeline itself rather than left to manual habit. If a summary is ever imprecise, the source is one tap away for anyone who wants to verify it — the same discipline I'd want in place for any AI system making a claim on someone's behalf.

The full translation lives on Telegra.ph. A summary tells you whether you care; it doesn't let you actually read the thing. Publishing the complete Farsi translation to Telegra.ph gives the channel post a link that opens as Instant View — the full article, in Farsi, in-app, for readers who would otherwise hit an English paywall or a blocked domain.

n8n over a hand-rolled cron script. A visual workflow keeps the pipeline legible and extensible — adding a topic or a second delivery channel means adding a node, not re-reading old scheduling code. The three newer workflows are generated from checked-in build scripts rather than edited on the canvas, so a deploy can't be quietly overwritten by a manual tweak.

De-duplication by design, not by luck. The source topics overlap more than you'd expect — a business story and a tech story often cover the same underlying event. Every candidate is checked against the normalized history of what has already gone out, and the history is only written once a message has actually shipped, so a failure mid-run can't blacklist an article that was never published.

Errors that announce themselves. An automation nobody watches fails quietly, and the first sign is usually a reader asking why the channel went dead. Every content workflow names the error-monitor workflow as its handler, so a failed run sends a direct message the same minute.

The Outcome

• Fully automated — credited, translated, Farsi posts across up to five slots a day, with zero manual work.
• Hours saved each week that used to go into reading, summarizing, and translating by hand.
• Consistent without being noisy — the channel never repeats itself, and stays quiet rather than padding a thin day.
• Self-reporting — a failed run raises a direct message instead of a silent gap.

Tech Stack

• AI: Google Gemini Flash — scoring, summarization, categorization and full-article translation
• Orchestration: n8n (self-hosted, version-pinned), with workflows generated from checked-in build scripts
• Delivery: Telegram Bot API via n8n's Telegram node; Telegra.ph for Instant View full translations
• Sources: RSS across tech, psychology, business and world news, each set to degrade rather than fail
• Infrastructure: Docker on a Frankfurt VPS, not exposed publicly — administered over an SSH tunnel

See it running — join Farasu on Telegram →