Every application, one screen, nothing on anyone's laptop.
This is the hub view your team works from. The engine drafts, checks and files; people confirm designations, handle SAQA and sign. Staff see this dashboard, never the machinery behind it.
1,187
applications this month
96.4%
passed every automated check first time
31
flagged for human review
11 min
median intake-to-draft time
Figures are illustrative, shaped on the roughly 1,200 applications per month discussed for the Huawei account.
In the pipeline now
Applicant
Type
Stage
Risk
Sipho Radebe
General work
Letter drafted, awaiting signature
low
Chen Jialing
General work
Document cross-check
medium
Ravi Pillay
Critical skills
Qualification match review
low
Wang Haoran
ICT
Skills transfer plan, year 2 draft
low
Naledi Khumalo
General work
SAQA evaluation in progress
human step
Li Wenjun
General work
Repeat applicant, prior file matched
medium
Repeat applicants are matched on passport, email and mobile before anything is drafted. The previous system fell over exactly here; this one treats a returning applicant as a first-class case.
Regulatory watchdog
Scraped daily at 05:45 from Home Affairs, the embassies you file with, and VFS. Changes land here before they land on you.
DHA16 JulCritical skills list, no change detected
VFS16 JulMumbai centre: biometrics slots moved to 48h lead time
Embassy15 JulUAE: attestation fee revision effective 1 Aug
Pick a sample applicant, confirm the designation, and watch the engine do what your team does by hand today: read the CV, reverse-engineer a job description that matches the experience, cross-check the documents, and draft the letter. This demo runs on invented data, right here in the page.
What the engine does
Read the CV
Pulls roles, dates, systems and qualifications into structured fields.
Reverse-engineer the job description
Builds a JD for the confirmed designation that matches this CV's actual experience, so the selection question always has an answer.
Cross-check the file
CV versus visa copies and passport: positions, dates, spellings. Mismatches get fixed before Home Affairs finds them.
Assess and tier the risk
Low, medium or high, with the reason written down. High risk never proceeds on its own.
Draft the letter set
One master template per letter type. A rule change is made once and lands everywhere.
A qualified human signs
Always. The engine prepares and proves; it never submits to a government.
Illustrative draft
The deductive check
It reads the file the way your best person does.
Every application is cross-referenced against the applicant's own documents before a letter exists. Here is a real shape of problem, on invented data: the CV and the previous work visa disagree.
From the CV
NameChen Jialing
PassportEA4471902
Role, 2023 to 2025Network Engineer II
Start date at Huawei SA03 / 2023
QualificationBEng Telecommunications
From the visa copy
NameChen Jialing
PassportEA4471902
Role on permitNetwork Technician
Permit valid from07 / 2023
ConditionEmployer-specific
Engine finding, tiered medium: the CV says Network Engineer II from March 2023; the permit says Network Technician from July 2023. Recommendation: align the CV to the permit for the overlap period and carry the promotion as an internal role change, with the permit as the anchor document. Fixed in the draft, flagged in the audit trail, shown to your reviewer before signature.
This is the check that answers "how did you select this person for this position" before it gets asked.
Compliance and SAQA
Honest about what stays human.
Some steps must not be automated: government logins, sworn translation, signatures. The engine wraps them instead: it prepares everything up to the step, tracks it while a person does it, and picks up the moment it is done. Nothing waits in someone's head.
SAQA evaluation tracker
Applicant
Step
Owner
State
Naledi Khumalo
Applicant email created, Nash-controlled
engine
done, 09 Jul
Naledi Khumalo
SAQA registration and document upload
human
done, 10 Jul
Naledi Khumalo
OTP handling on SAQA logins
human
as needed
Naledi Khumalo
Sworn translation of qualification
human
in progress
Naledi Khumalo
SAQA report filed into the application
engine
waiting on step above
Naledi Khumalo
Letters drafted using the SAQA outcome
engine
queued
The tracker chases the step, not the person: anything human that sits longer than its service window shows up in the morning summary.
Why this earns the percentage: at volume, nobody re-reads 1,200 letters. What protects the firm is a system where every figure is checked deterministically, every judgment is tiered and logged, and every letter that goes out carries a trail showing exactly why it says what it says.
End to end
The workflow, with the human steps named.
Accuracy at volume does not come from trusting a model. It comes from a deterministic core, bounded judgment, and gates that fail closed. This is the whole path an application takes.
01
Intake
Client uploads through a guided link: one question at a time, documents at the end. Everything lands in the case file with a timestamp; the audit trail starts here.
automated
02
Duplicate and repeat matching
Passport, email and mobile checked against every prior application. A returning applicant continues their file rather than colliding with it.
automated
03
Designation confirmed
Nash confirms the position with the client before any drafting. If the position sits on the critical skills list, the engine proposes three workable alternatives.
human
04
Extraction and verification
Fields read off the documents, then checked by more than one independent model; disagreement means a person looks. Passport validity, dates and spellings checked deterministically.
automated
05
Reverse-engineered job description
The general work visa hero step: a JD built to match the CV's real experience for the confirmed designation.
automated
06
Document cross-check and risk tier
CV versus visa copies versus passport. Fixes proposed, everything logged, low, medium or high tier attached. High risk stops the line.
automated
07
SAQA, translations, attestations
Prepared and tracked by the engine, performed by people: registration, OTPs, sworn translation. The tracker chases anything that stalls.
human
08
Letter set drafted
One master template per letter type, per-client signatory blocks pulled automatically, correct every time. A rule change is one edit, everywhere at once.
automated
09
Quality gates
Every figure traced to a source document, every name spelled identically across the set, template version recorded. Any failure holds the letter; gates fail closed.
automated
10
Review and signature
A qualified person reviews the flagged minority and signs everything. The engine never signs and never submits to a government portal.
human
11
Filed, tracked, learned from
Outcome recorded against the application. Every exception a reviewer catches becomes a new deterministic check. The system gets stricter every month it runs.
automated
The design rule behind it: anything that can be checked by computation is checked by computation; judgment is bounded and logged; the irreversible steps (government submissions, signatures, money) stay with named people. That is how you run 1,200 a month with no comebacks.
Analytics
The numbers Rayne sees without asking anyone.
Volumes, timings, exception rates and the money, computed from the audit trail itself, updated continuously. This month, on illustrative data:
1,187
letter sets produced
2.6%
exception rate, trending down
R 41.8k
estimated staff hours returned, at cost
4
regulatory changes caught by the watchdog
Time per letter type: manual today versus engine
General work motivation
95 min manual
11 min engine
Skills transfer plan
4 to 6 h manual
38 min engine
Critical skills assessment
80 min manual
14 min engine
Engine times include the human confirmation steps they wrap. Manual times as described in our meetings; to be measured properly in the side-by-side trial weeks.
What a month costs to run
Line
Basis
Monthly
Document reading and judgment steps
~1,200 applications, per-document model calls
under $10
Letter generation
template fill, effectively free
~$0
Server, one country
dedicated VPS, data stays in-country
$20 to 60
The token fear from our first call, answered with arithmetic: generation is deterministic, models are only paid for judgment moments, and the whole machine runs for less than one manual letter's labour.