Side A
Risk
Read both sides at once
Bespoke risk analytics
One month of monitoring for a single client draws on more than 267,000 posts and articles.
Risk analytics · Rapid platforms
Frontier Analytics
Read the risk, build the platform
We make two things. Side A is bespoke risk analytics, built from your own records and open sources. Side B is rapid platforms: hyper-bespoke systems where data meets operations, prototyped on your own data in weeks.
Formerly
VoxCroft Analytics
Worked in
EnglishAfrikaansisiXhosaisiZulu
HausaAmharicIlocanoHokkien
JavaneseSpanishHaitian CreoleSomali
HindiChineseVietnameseLingala
The short version
Two sides, one engine
Frontier Analytics makes two things, and the second grew out of the first.
The first is bespoke risk analytics. We join what a company already knows about its own risk, in incident reports, site logs, and security registers, to what the outside world is saying in news, social media, and local languages. Then we build the platform that puts it in front of the people who act on it, in the way they already work.
The second is rapid platforms. Building risk platforms taught us how to turn years of unread records into something a team can run on. Now we do that wherever data meets operations: deal flow, programmes, due diligence, and reporting.
-
Your own records and open sources, read together, and delivered as a platform cut to your team's workflow.
Play side A -
Side B
Platforms
Prototyped on your own data
Rapid platforms
One archive audit read 1,021 documents without collecting any new data.
Hyper-bespoke platforms where data meets operations, prototyped on your real data in weeks and built out in phases.
Play side B
One engine underneath both
Machines handle the volume, people handle the judgement, and every finding traces back to its source. That holds on both sides of the record.
Side A
Risk
Read both sides at once
Bespoke risk analytics. We join your own records to the open sources around your sites, in the languages people actually use, and build the platform your team acts from.
Three stops
on this side
Side A · Where the risk hides
Four ways it goes missing
The risk that matters most sits in two places at once: inside your own business, and out in the open where your tools don't reach.
-
Language
It isn't in English
People talk about a plant, a mine, or a policy in the languages they actually speak. Our monitoring for one carmaker runs in English, Afrikaans, isiXhosa, and isiZulu for exactly that reason.
-
Offline
It isn't online
Some of the most important views never reach a search engine. They live in community meetings, local radio, and conversations only a person on the ground hears.
-
Inside
It's already in your own records
Incident reports, site logs, and security registers pile up as text nobody has read side by side. The business already holds part of the answer but can't see it.
-
Volume
267,000+
There's too much of it
One month of monitoring for a single client now draws on a sample of more than 267,000 posts and articles. No team reads that unaided.
General-purpose AI tools don't fix this on their own. They learn mostly from English, online, well-structured data, so they tend to sound most confident where they know least. When our team trained translation models with local speakers, our Hausa model was 15.6% more accurate than a commercial one.
So we read both sides at once: your own records and the open sources around your sites, in the languages people actually use.
Side A · Cut to your workflow
Built around the desk
We don't sell a dashboard and ask your team to move into it. We build the risk platform around the desk, the shift, and the channel your people already use.
Three inputs
Inside
Your records
Incident reports, site logs, security registers
Outside
Open sources
News, social media, named local accounts
On the ground
Local experts
Checks and field reports
One platform
The platform
Built to fit
- Collect
- Translate
- Classify
- Score
- Trace to source
People make the call
Into your workflow
When a line is crossed
Flash alerts
By WhatsApp, SMS, or email
Every month
Risk assessment
With a tracked risk score
Always on
One view
Inside and outside, every item sourced
Your records tell you what has already happened at your sites. Open sources tell you what people around them are saying now, in their own languages. Local experts check what monitoring picks up and hear what never goes online. The platform joins the three, and analysts decide what it means before anything reaches your team.
-
Into the channels you already watch
Flash alerts arrive by WhatsApp, SMS, or email, on the shift pattern your team works, each one linked to the source it came from.
-
Judged against your own baseline
Each site is measured against what normal looks like there, so a monthly risk score means something to the people who read it.
-
Your records, made readable
Incident reports and site logs become structured data, joined to the open-source picture in one place, with views built around the decisions your team actually makes.
Your data stays where your rules say
Internal records are the most sensitive part of any risk picture, so we design around them. We host in South African data centres or install on your own hardware, strip personal information before anything is read, and give reviewers snippets rather than whole records. When your rules require it, the data can stay in the country, or in the building.
Who it's for
Security, risk, and corporate affairs teams at companies whose sites, people, or supply chains are exposed to unrest, crime, or reputational risk.
Side A · On the record
Risk on the record
Live work and past work on the risk side. Clients are described here rather than named.
267,000+
posts and articles in one month's monitoring sample
A carmaker's monthly assessment
3,000+
alerts logged for one client since July 2025
The same retainer
479
internal incident reports turned into structured data
A mining group's own records
1,889,963
translated sentence pairs behind the language reach
Eight African languages
Side ARisk platforms
-
A1
A global carmaker's South African arm
Since July 2025
What we did
Monitoring of X, news, and named local accounts in English, Afrikaans, isiXhosa, and isiZulu around its plants and dealerships; flash alerts; and a monthly risk assessment series since March 2026.
LiveWhat it showed
The monthly sample grew from 27,688 items in the first assessment to more than 267,000 in the sixth. More than 3,000 alerts logged since July 2025.
-
A2
A major South African mining group
2025 to 2026
What we did
A brief on public attitudes to illegal mining; 479 internal incident reports turned into structured data; and weekly field reports and community briefs near its operations.
ExploratoryWhat it showed
Early signs that low-level incidents build up in the hours before serious ones. Confirming it needs 12 to 24 months of data.
-
A3
A global health foundation
2021 to 2024
What we did
Security-threat and misinformation monitoring, starting in Ethiopia and growing to five countries.
DeliveredWhat it showed
Alerts by SMS, WhatsApp, and email, curated by an analyst, plus monthly reports.
-
A4
An international consultancy
2021 to 2024
What we did
Standing collection on attitudes to a Gulf state across more than 75 countries, with fortnightly reports.
DeliveredWhat it showed
Crowd-verified datasets coded for sentiment and themes, plus bot and disinformation analysis.
UnderneathThe engine
-
E1
Our own alert pipeline
2025 to 2026
What we did
Collectors that pull South African headlines every 12 hours, an LLM that applies nine alert thresholds and ten categories, and an automated monthly risk workflow.
LiveWhat it showed
About two-thirds of alerts since late June 2026 drafted by the agent, with analysts doing the rest.
-
E2
Our protest and conflict forecasting model
2026
What we did
Tested on 48 countries, measuring risk as a change from each country's own baseline.
In researchWhat it showed
Adding local-language text signals improved how well it ranks risk (PR-AUC up 0.048 to 0.070). It is not yet in operational use, so we present it as research.
GroundworkLanguage reach
-
L1
A government language-technology programme
2022 to 2024
What we did
Crowd-built translation data for eight African languages, collected and checked by our network of local linguists.
DeliveredWhat it showed
1,889,963 sentence pairs. Amharic and Hausa models scored 36 and 37 on SacreBLEU, and the Hausa model was 15.6% more accurate than a commercial one.
-
L2
A government research programme in the Asia-Pacific
2024 to 2025
What we did
Text and audio collection in Ilocano, Hokkien, and Javanese.
DeliveredWhat it showed
About US$2 million of work over 18 months, alongside two major technology and research partners.
-
L3
A medical interpreting platform
Late 2025
What we did
Post-editing of machine translation and medical speech recordings in Spanish, Haitian Creole, Somali, Hindi, Chinese, Vietnamese, and Lingala.
DeliveredWhat it showed
Every job completed and delivered by December 2025.
Between the sides
Five rules we keep
Five rules came out of the risk work. They run through everything on both sides of the record.
Collect what matters, not what is easiest to collect
We go to the source, even when that means a new language, a new collection method, a person on the ground, or a shared drive nobody has opened in years.
Machines for volume, people for judgement
AI reads, translates, sorts, and drafts. Analysts, local experts, and your own team decide what it means. Since late June 2026, about two-thirds of one client's alerts were drafted by an AI agent; analysts wrote the rest and reworked the agent's drafts where needed.
Every finding traces to a source
Our reports are footnoted, and the platforms we build link each fact to the document it came from. If we can't show where something came from, we don't say it.
Judge each place against itself
A protest count means little until you know what normal looks like there. Our forecasting research measures risk as a change from each country's own baseline, and our platforms measure each site, programme, or portfolio against its own history.
Build things that run every day
A one-off report goes stale. We set up monitoring, agents, and data stores that keep working, and we stay on to run them or hand them to your team.
Side B
Platforms
Prototyped on your own data
Rapid platforms. Risk platforms meant turning years of unread reports into data a team could act on. Clients started asking for the same thing everywhere else in the business.
Three stops
on this side
Side B · Where data meets operations
Four shapes it takes
Rapid platforms are hyper-bespoke: built for one organisation, around one workflow, on that organisation's own data.
They sit where data meets operations, in the places where the real work runs on documents, inboxes, and group chats, and nobody can see the whole picture.
-
Deal flow
One pipeline for every lead
Inbound, outbound, and network leads land in one queue, get enriched and scored against your criteria, and move through your gates to an investment committee memo. The reason for every decision is kept, so it sharpens the next one.
-
Programmes
A programme that runs off one record
Intake, scoring, coaching notes, compliance chasing, and sponsor reports, with founders answering over WhatsApp and your team staying on the board it already uses.
-
Data rooms
A data room with two sides
Companies fill it from guided templates and see how complete it is. Investors get invite-only access, with a log of who opened what.
-
Inbox and chat
An agent in the group chat
It joins WhatsApp groups and email threads, files what is shared under one naming structure, records decisions as they are made, and answers "send me the latest management accounts" with the source attached.
Work stays where the team already does it
Nobody is asked to move into a new tool. The platform drafts, checks, and chases behind the inbox and the group chat; people review, decide, and sign off.
Rules before models
Most of an organisation's know-how is a rubric, a checklist, or a set of gates, and those are configuration rather than AI. We write them down first, and add models only where there is enough data to trust them.
Side B · How it runs
Prototype first
Every engagement starts small, puts something working in your hands early, and earns its next step.
-
01
Audit what you already hold
A short, fixed-price read of your drives, inboxes, and chat exports. No new data collected. It ends in numbered findings that every later piece of work traces back to.
-
02
Prototype on your own data
A first working version in weeks, on your real records, so you judge the thing itself rather than a slide about it.
-
03
Build in phases
Each phase ends with something working that you keep, and a decision point where you can stop. Work is priced per delivery day and quoted firm for each phase once it is scoped.
-
04
Run it, or run it yourself
We run it for a monthly fee you can cancel on 30 days' notice, or hand it over so your team runs it without us in the room. Hosted by us, or on your own hardware so the data never leaves your premises.
| Stop | Takes | You leave with | Fare |
|---|---|---|---|
| Audit | Days | Findings | Fixed price |
| Prototype | Weeks | First version | Firm quote |
| Build | Phases | What you keep | Per phase |
| Run | Monthly | Live platform | 30 days out |
Risk engagements on Side A start the same way: a diagnostic, then phases, then a monthly fee.
Who it's for: organisations whose operations run on documents, inboxes, and group chats. Investment teams and their portfolio companies, programme operators, and mid-market firms who want their own information working for them.
Side B · On the record
Platforms on the record
What one archive audit turned up, and the builds on this side so far. Clients are described here rather than named.
1,021
documents read in one programme archive
2014 to 2023
96
companies with a score on file
The same archive
4
scoring instruments compared for the first time
The same archive
0
new data collected for the audit
Every number came from files already on the drive
Side BRapid platforms
-
B1
A pan-African enterprise-development programme
2026
What we did
A fixed-price audit of the programme's archive: 1,021 documents from 2014 to 2023, 96 companies scored, and four scoring instruments compared side by side for the first time.
Audit doneWhat it showed
Every number came from files already on the drive. The numbered findings set the build plan: a short pilot with founders answering over WhatsApp, then programme automation on the team's own board and a two-sided data room.
-
B2
A social platform
2026
What we did
A technical risk review, a UX audit, and ten fixed-fee fixes to the platform.
DeliveredWhat it showed
Onboarding, connections, email personalisation, and photo handling fixed.
-
B3
A pharmaceutical software company
2026
What we did
An agent-assisted CI/CD and GitOps platform for the rewrite of their product.
DeliveredWhat it showed
Delivered in four milestones with a specialist partner.
The long version · about two minutes
How one side grew the other
Our team comes from open-source intelligence and risk analysis: years of work for companies, foundations, and international programmes, first as VoxCroft Analytics and now as Frontier. That work kept taking us to places where the usual data runs out: languages commercial AI handles badly, communities that never show up in a search, and organisations whose most useful records had never been analysed.
So we built what was missing. We trained a network of local linguists and experts, and with them produced nearly 1.9 million translated sentence pairs and models for eight African languages. Then we built monitoring that reads South African social media in four languages and drafts alerts with AI, while analysts make the calls. A major manufacturer uses it to watch for risk around its plants and dealerships, with a monthly assessment drawn from more than 267,000 posts and articles.
The turn came when clients handed us their own records. A mining group's incident reports became structured data that showed patterns nobody had seen. The engine that read the outside world could read the inside of a business too, and once it could do that, it could run the work around it.
That became the second side of the record. A programme's archive of more than a thousand documents, read and scored. Deal flow, data rooms, and group chats turned into platforms built for one organisation, prototyped on its own data in weeks.
That's why we're called Frontier. The word has two meanings for us. It's the place: frontier markets, where data is thin and the stakes are high. And it's the technology: current machine learning and AI, where we know from experience what works and what still breaks.
It's where the map runs out
Last call
Book a diagnostic
Both sides start the same way: a short, fixed-price diagnostic of your data and the question you need answered.
Tell us the risk you're watching or the work that's stuck, and where you think the information is. We'll tell you whether we can reach it, and what it would take.
Or write to