SONAR monitors online news, video and social media for the topics you set, filters out what is off-topic, measures public stance toward your organisation, and scores the risk of issues that start to rise. Your team reads it in one place, refreshed every 30 minutes.
Public issues rarely start on the front page. They start in comment sections, short videos and Telegram channels, and then the press picks them up. SONAR reads all of those channels in one database, with relevance rules written for your organisation, and tells your team when an issue's volume rises outside its usual pattern.
01Product
From the morning scan to the report for leadership. Every image below is an unedited screenshot of the demo project.
01 · Radar
Every two hours the latest conversation is grouped into issues and scored for risk from 0 to 100 on five drivers: volume, velocity, sentiment, spread and the share of copy-paste text. An 18-hour forecast is shown with its range.
What your team can do
02 · Early warning
An issue is called a spike only when its hourly peak passes a statistical test against that issue's own baseline. The curve is split into trigger, amplification, peak, decay and follow-up waves, with the first accounts and outlets that carried it.
What your team can do
03 · Mention explorer
Mention explorer lists the articles, videos, posts and comments behind every chart. Search takes operators: quoted phrases, OR, exclusion with a minus sign, site: and author:.
What your team can do
04 · Sentiment
A project can use tone mode or stance toward one target, such as your ministry or company. In stance mode, criticism your organisation makes of others counts as neutral. The page always states what share of labels comes from the language model and what share from the lexicon.
What your team can do
05 · Network
The network graph links people, organisations, terms, hashtags and accounts that appear in the same mention. The Entities page ranks the most-named names with their sentiment and channels.
What your team can do
06 · Reports
Build a report from dashboard cards, save the layout, then export to PDF or Word. The numbers in the report match the dashboard, and the AI narrative is optional.
What your team can do
07 · Campaign Decoder
Campaign Decoder breaks a TikTok and Instagram creator campaign into formats, a second-by-second script and a performance map. It also judges whether the wave of videos arose on its own or follows a patterned brief.
What your team can do
02Decoder
Campaign Decoder reads creator campaigns on TikTok and Instagram, whether a brand campaign or a wave of videos about a public issue. Every video is sorted into a format, its script is broken down by the second, and its performance is compared with the campaign median.
01
AI groups videos into formats discovered from the campaign itself, then breaks each script into hook, brand or issue appearance, method, benefit and CTA, with timestamps.
02
Each video is placed by views and likes against the campaign median and against its creator's usual numbers, so a video that wins on format is separated from one that wins on a large account.
03
Eleven signals, from posting times to near-identical captions, are combined into a 0 to 100 score, and AI reconstructs the probable brief together with the evidence against it.
From the database
1,597 videos from 7 decoded test campaigns. Campaigns with sample data are not counted, and no campaign or account names are shown here. Last decoded 3 Nov 2026, 14:45 WIB.
Used by
03Solutions
Three ways organisations use SONAR. Each flow uses pages that exist in the product today.
01
Public relations, communication bureaus, control rooms
A policy is announced on Monday morning. The press office needs to know whether opposition is growing, and where.
02
Secretariats, parliamentary groups, advocacy teams
An organisation wants to know how the public judges its position on a bill, separately from the criticism it makes of others.
03
Corporate communications, customer care
A price rise lands this week. The team wants to find complaints in video comment sections before the press quotes them.
04Platform
Every box below runs today. Blue lines mark where numbers are checked before they are shown; circles mark where your analysts can correct the machine.
01
Collectors run every 30 minutes for every project: news feeds, YouTube, Telegram, Bluesky, Google News, and data providers for TikTok, Instagram, X, Facebook and Threads.
02
Each item is scored with the project's relevance rules: anchor phrases, acronyms, names, context words and negative words. Items near the threshold are judged by a language model.
03
A language model labels relevance, tone, stance toward the project's target, emotion, categories and a summary. Daily counts per channel and sentiment refresh every 15 minutes.
04
Early warning and Radar flag rising issues. Response plans, key messages, per-channel drafts and reports are built from the same issue data.
Before an issue summary, early-warning narrative, draft, report or AI answer is shown, any sentence that cites a number not found in its source data is removed.
05Live data
The charts in this section are recomputed every hour from the demo project National Political Issues. Each one states how it is computed.
Relevant mentions per day, last 90 full days
Relevant, non-duplicate mentions per channel, last 30 full days in Jakarta time. Comments count as mentions.
People and organisations from entity recognition, counted per mention. Places, products and political parties are not shown on this public page.
National Political Issues
tone mode
negative 31%neutral 57%positive 12%
99% of 92,546 mentions labelled
Only mentions labelled by the language model or an analyst. 99% of relevant mentions from the last 30 days are labelled so far.
06Methodology
We tested the automatic labels against 508 mentions labelled by a person. The results are published as they are.
recall 88.7%, 508 mentions
232 relevant mentions
40 mentions, tone-mode project
Right now 99% of the 92,546 relevant mentions from the last 30 days are labelled by the language model or an analyst. Across the whole database the figure is 100%; the rest carry lexicon labels and are marked on the Sentiment page.
| Measured | Result | Sample | Note |
|---|---|---|---|
| Sentiment, all relevant mentions | 67.8% | Sample: 311 | accuracy |
| Tone-mode project | 72.8% | Sample: 232 | National Political Issues |
| Stance-mode project (demo) | 53.2% | Sample: 79 | about 72% when scored with codebook v2 |
| Keyword lexicon | 65.3% | Sample: 311 | fallback until model labels arrive |
| Relevance, precision | 87.9% | Sample: 508 | recall 88.7% |
| Emotion, precision | 37.7% | Sample: 508 | read as an indication, not a measurement |
| Spot check of live labels | 92.5% / 87.5% | Sample: 40 + 40 | tone project / stance project |
Measured on 24 October 2026 with the previous model version. A re-run with the current model has not been done yet.
Rules are written per project. This excerpt is from the demo project: "MK" only counts when capitalised, sports and foreign politics score negative, and items between the two thresholds are checked by a language model.
{
"anchors": [
{ "p": "presiden|wapres|istana|kabinet|…", "w": 3 },
{ "p": "MK", "w": 3, "case": true },
…
],
"negatives": [
{ "p": "piala dunia|liga 1|sepak bola|…", "w": -2 },
{ "p": "bundestag|midterms|white house|…", "w": -4 }
],
"threshold": { "relevant": 3, "irrelevant": 0 },
"sentiment": { "mode": "tone" }
}07Security
SONAR is a web application with a PostgreSQL database. Everything inside the dashed line runs in one place: your server, or a Grimmia-managed VPS in Jakarta.
On your organisation's server or data centre, installed with the Grimmia team, or on a Grimmia-managed VPS in Jakarta. With the second option each client gets separate projects and sees only its own.
Each role gets read or read-write access per page. Access is recorded in the activity log, and API tokens can be revoked at any time.
AI sentences that cite numbers not found in the source data are removed before they are shown. The Brief paragraph on the dashboard uses no language model at all.
Only mention text, at most 800 characters per item, goes to the language-model provider through the gateway for labelling and summaries.
Set in the contract. When the contract ends, project data is exported for you and then deleted from Grimmia-managed servers.
Grimmia does not train AI models on your data.
08Sources
The source registry is checked automatically every day. Feeds or channels that fail are not used until they recover.
253
news feeds (RSS)
92
YouTube channels
33
Telegram channels
6
social platforms
Television, radio and print are not monitored yet, except what is uploaded to YouTube.
Last checked 25 Oct 2026, 13:13 WIB
National 60
Antara 10, Republika 8, Detik 6, Liputan6 5, Sindonews 5, CNN Indonesia 4, Tempo 3, Kumparan 2, Akurat, Alinea, Disway, iNews, Jawa Pos, JPNN, Jurnas, Magdalene, Medcom, Media Indonesia, Okezone News, Project Multatuli, RMOL, Tagar, The Conversation ID, Viva, VOI
Business 25
CNBC Indonesia 8, Detik Finance 4, Kontan 2, Liputan6 2, CNN Indonesia Ekonomi, IDX Channel, Investor Daily, Katadata, Kumparan Bisnis, Okezone Economy, Republika Ekonomi, Sindonews Ekbis, Warta Ekonomi
Government 18
Bawaslu RI, Pemprov Jateng, Kemenko PMK, Kementerian Pertahanan, Kementerian Perdagangan, Badan Narkotika Nasional, Lembaga Penjamin Simpanan, Pemprov Sulteng, Pemprov Aceh, Tribratanews Polda Gorontalo, Tribratanews Polda Aceh, PPID Pemprov Jateng, Pemkot Samarinda, Pemkot Malang, Pemkot Banjarmasin, Pemkab Bantul, Pemkab Karawang, Pemkab Gunungkidul
Parliament 5
Fraksi Gerindra, Fraksi PKB, Parlementaria DPR, DPRD Jawa Tengah and 1 more
Parties 5
Partai NasDem, Golkarpedia (Partai Golkar), Partai Gerindra, PKB and 1 more
International, Indonesian-language 4
Antara English, BBC Indonesia, DW Indonesia, Anadolu Indonesia
Fact-checking 1
Liputan6 Cek Fakta
Regional feeds by province 135
Jawa Timur 23, Jawa Tengah 11, Bali 7, Jawa Barat 7, Sumatera Selatan 6, Sumatera Utara 5, Bengkulu 4, DI Yogyakarta 4, Kalimantan Timur 4, Kalimantan Utara 4, Nusa Tenggara Barat 4, Papua 4, Riau 4, Sulawesi Selatan 4, Aceh 3, Banten 3, Gorontalo 3, Kalimantan Barat 3, Kalimantan Selatan 3, Kepulauan Riau 3, Lampung 3, Nusa Tenggara Timur 3, Sulawesi Utara 3, Sumatera Barat 3, Jambi 2, Kalimantan Tengah 2, Kepulauan Bangka Belitung 2, Sulawesi Tengah 2, DKI Jakarta, Maluku, Maluku Utara, Papua Barat, Papua Tengah, Sulawesi Tenggara
National 29
KompasTV, CNN Indonesia, tvOneNews, Metro TV, Kompas.com, detikcom, Tribunnews, Official iNews, Seputar iNews RCTI, BeritaSatu, Narasi Newsroom, Tempodotco, TV Tempo, SINDOnews, Liputan6, TVRI Nasional, RRI NET Official, Okezone, Antara TV Indonesia, Republika Official, JawaPos TV, kumparan, IDN Times, VIVA.co.id, Media Indonesia, Tirto ID, Suara.com, Merdeka.com, Pikiran Rakyat
Regional 13
Kompas TV Jawa Timur, Kompas TV Jawa Barat, Kompas TV Jawa Tengah, Kompas TV Dewata (Bali), Kompas TV Lampung, Kompas TV Aceh, Kompas TV Makassar, Tribun Timur, TribunJatim Official, Tribun Jabar Video, TribunJakarta Official, Tribun Medan TV, Serambinews
Business 5
CNBC Indonesia, Katadata Indonesia, Kontan TV, Bisnis.com, IDX Channel
Government 30
Sekretariat Presiden, Sekretariat Kabinet RI, Kementerian Sekretariat Negara RI, Kementerian Keuangan RI, Kemkomdigi TV, Kemendagri RI, MoFA Indonesia (Kemlu), Kementerian Kesehatan RI, Kemhan RI, Kemenko Polkam RI, Kemenag RI, Kementerian Hukum RI, Info BMKG, BNPB Indonesia, KPU RI, Bawaslu RI, KPK RI, Divisi Humas Polri, Polda Metro Jaya, Mahkamah Konstitusi RI, Mahkamah Agung RI, Kejaksaan RI, TNI AD, Info BNN RI, Bank Indonesia Channel, Otoritas Jasa Keuangan, Pemprov DKI Jakarta, Humas Jabar, Pemprov Jawa Tengah, Kominfo Jatim
Parliament 4
DPR RI, TVR Parlemen, MPR RI and 1 more
Parties 9
Gerindra TV, PDI Perjuangan, DPP Partai Golkar, DPP PKB, NasDem TV, Partai Demokrat, PAN TV, Partai Solidaritas Indonesia and 1 more
International, Indonesian-language 2
BBC News Indonesia, DW Indonesia
Telegram channels 33
18 official; 15 unofficial channels (mirrors or aggregators) are flagged and filtered more strictly.
social platforms
The sites below offered no feed we could fetch directly when checked on 25 Oct 2026. Their articles arrive through Google News keyword search, and some through their official YouTube channels.
09FAQ
No. SONAR collects search results for the keywords, accounts and links you set, within each platform's quota. The result is a consistent day-to-day sample of the conversation, enough to read trends and issues, but not a count of every post.
On a 508-mention test set, language-model labels were right 67.8% of the time for relevant mentions and 72.8% in the tone-mode project. Stance mode is still at 53.2% against the old test set. Details are in the accuracy section above and on the methodology page.
Yes. Analysts can change sentiment or mark a mention off-topic in Mention explorer. Manual corrections always win and are never overwritten.
Collectors run every 30 minutes, language-model labels are added every 10 minutes, and Radar is recomputed every two hours. The statistics on this public page refresh hourly.
Yes. Keywords, accounts, Telegram channels and post links are set per project. New RSS feeds and YouTube channels are added to the registry and checked automatically before use.
On your organisation's server, or on a Grimmia-managed VPS in a Jakarta data centre. Mention text is sent to a language-model provider for labelling; Grimmia does not use it to train models.
SONAR only collects publicly accessible content. Private accounts and closed groups are not collected. Individual posts are visible only to that project's users, never on public pages like this one.
No. Campaign Decoder gives a 0 to 100 structure score from eleven signals, such as posts close together in time, the same hashtag bundle and near-identical captions. A high score means a strong pattern, not proof of payment. Every result lists the counter-evidence and what a person should check, such as paid-partnership labels and account history.
Yes. Admins can create API tokens to read analytics data, and mentions can be exported to CSV or Excel. Tokens can be revoked at any time.
10Contact
Tell us what you want to monitor. We prepare a demo with data on that topic, then put together a quote.
A quote depends on
Farid Mahardika, Grimmia Intelligence