okxproject
All work
Live

SocialRadar Alert

Same-day alerts on what was just published about a brand, sent to the WhatsApp group its team already uses.

Status
Live since August 2026
Sources
YouTube, Instagram, listening alerts
Delivery
WhatsApp groups and email
Scale
827 accounts, 12 groups

How an alert gets through

Simulation with sample data

Sources

YouTube channels705
Instagram accounts122
Spoken mentions304 phrases
Listening alerts6,020 read

The brand's rule

KeywordDoes it name the brand or its topics?
View thresholdHas it been seen enough to matter?
Topic filterIs it a topic the team asked to skip?
Brand team groupWhatsApp
    Checked
    0
    Held back by the rule
    0
    Delivered to the group
    0

    Why it exists

    A brand team should not have to walk through three platforms and a dashboard to learn what was said about them today. SocialRadar Alert brings it to the place they already look, while the content is still fresh.

    • Speed. Content shows up in the group the day it is published. It does not wait for the month-end report.
    • Selectivity. Only content that fits the brand's own rule goes out. A rule is made of keywords, a view threshold and topics to skip.
    • One channel. Account tracking, search, spoken mentions and listening alerts all land in the same group, in the same layout.

    How it works

    1. Watch

      705 YouTube channels and 122 Instagram accounts are checked in parallel. Channels outside the list are found by search every 20 minutes.

    2. Listen

      The spoken track of each video is searched every hour for 304 phrases covering 101 concepts. A brand named out loud is caught even when the title says nothing.

    3. Filter

      Each item passes the brand's rule: keyword, view threshold, language threshold, topic filter. Listening alerts also pass an AI relevance check.

    4. Deliver

      What passes is written to the brand's WhatsApp group. Where a team prefers, the same alert arrives by email at the same moment.

    5. Report on itself

      Every hour the system posts one message on the health of all its services. Nothing goes live without approval.

    827
    accounts on the watch list
    6,020
    listening alerts read and routed so far
    304
    phrases searched for inside spoken video, every hour
    3,030
    automated tests run before any release

    From one tracker to twelve groups

    1. Started with Instagram account tracking.
    2. YouTube added.58 news channels, 135 accounts in total.
    3. Listening alerts routed into WhatsApp groups.Sorted by category for the teams that asked for it.
    4. AI relevance filter.Unrelated mentions are dropped before they reach a group.
    5. Parallel scanning.Seven workers instead of one. A round of 85 accounts used to take 25 minutes.
    6. Moved from a desktop to a server.The alert line was down for about 17 seconds.
    7. A 248-agent bug sweep.It found a caching race that had produced duplicate alerts. Fixed the same night.
    8. The watch list grew from 195 to 819 accounts.
    9. Spoken-mention search became a permanent service.
    10. A customer's feedback became a live rule the same day.After a rehearsal on 129 past alerts.

    How much the AI may do on its own

    Independence does not mean working without approval. Each step up the ladder is a separate decision.

    1. Read and measureFree
    2. Propose a change and rehearse it on past dataFree
    3. Reversible actions the customer never seesWith notice
    4. Anything that changes what the customer sees: code, rules, recipientsWith approval
    5. Test messages to a customer group, or AI-written text nobody asked forNever

    Where the AI sits

    Claude works beside us as a developer: it writes code and tests, measures on live data and prepares each release. At run time a small Claude model decides whether a listening alert is really about the brand.

    Bulk work stays on the cheapest model that passes a blind test. When 19 videos were scored without knowing which model wrote which summary, the smaller model scored 4.40 against 4.52 for the larger one, at one sixth of the cost. The larger model is kept for work that needs judgment: writing a rule, auditing it, turning feedback into a rule.

    What comes next

    • A new brand in one working day, opened from a panel with no code change.
    • TikTok as a source. The code and keywords are ready.
    • More than alerts: a daily digest, a weekly bulletin and a searchable archive.
    • Quality measured every week, brand by brand, so problems are seen before a customer reports them.

    Want something like this for your team?

    Tell us what you do by hand today. We will tell you what can be built.

    Email okan@okxproject.com