Strictly confidential

VFS Travel
Intelligence Engine

Accurate, sourced answers for every visa corridor, kept current automatically.
Prepared for

VFS Global

Prepared by

360 Labs

Technical and commercial proposal

September 2026

00Summary

One trusted answer for every visa corridor

We map where visa information lives, verify every claim against its source, and re-check only what changes. Travellers get answers read from that verified record, not generated on the fly.
The problem
39,800

Possible origin and destination corridors. The rules for each are split across VFS pages, mission sites, other operators and PDFs, and they change without notice.

What we built
99.6%

Of claims grounded in source evidence (510 of 512) in a working prototype across 31 corridors. 280 visa types tagged. Zero model calls when answering.

What we propose
2 months

To take the prototype to VFS’s 1,826 published corridors, then expand outward. $26,000 build. Running the VFS estate costs about $90 a year in infrastructure.

02 / 16
01The problem

One visa question, many sources

“What documents do I need for a student visa from India to Germany?” The answer is split across VFS pages, mission sites, other operators and PDFs, each owned by someone different.
4 of 31

Ownership is not obvious

In 4 of the 31 corridors we examined, the service is run by someone other than VFS. Finding the right source starts with knowing who is responsible.

212

The same thing has many names

We found 212 distinct visa-type names across sources. The same visa is described differently by every publisher.

95

Sources change quietly

We found 95 operational notices. Any one of them can change an answer without the answer changing with it.

Finding the information once is easy. The hard part is knowing which source is authoritative, reading it correctly, and keeping the answer true.

03 / 16
01The problem

At scale, it becomes an engineering problem

Each corridor needs six things to go right, and there are close to 40,000 corridors.
39,800

corridors

200 countries × 199 destinations

~400K

pages to process

Planning assumption: 10 per corridor

~16K

diplomatic posts

Upper bound of the source universe

1  Discover

Find the right sources for each corridor.

2  Access

Reach static pages, dynamic pages, APIs and PDFs.

3  Understand

Turn different formats into one record.

4  Verify

Prove every claim against its source.

5  Monitor

Detect the changes that matter.

6  Review

Put people where judgement is needed.

04 / 16
02The proof

The prototype works on real sources

We ran the full pipeline on 31 corridors (29 Schengen, the US and New Zealand), using 129 live VFS, government and operator sources and 65 PDFs.
What it showed
99.6%

of claims grounded in source evidence

510 of 512 claims were supported by the fetched source. The other 2 were rejected automatically.

280

visa types tagged by purpose

0

model calls when answering. Answers are read from the knowledge base.

Discover

Find and validate the relevant sources for each corridor.

Harmonise

Turn different page and document structures into one usable record.

Ground

Tie every claim to the exact source text that supports it.

Materialise

Store the approved result, ready to serve.

This proves technical feasibility on a limited set of corridors. Production-scale validation and wider coverage come next.

05 / 16
02The proof

Try the working prototype

Pick an origin and destination, then ask. Every answer carries a verbatim quote from the source it came from.
chancery.360labs.tech/sandbox Live Open in new tab

Live prototype

The sandbox could not be loaded here. It may be offline, or you may have no network connection.

Open chancery.360labs.tech/sandbox
06 / 16
03The system

How the system works

Two loops build and maintain a verified knowledge base. Answers are read straight from it, so serving is fast, cheap and traceable to source.
Sources

VFS websites

Country pages, visa information

Government sources

Ministries, embassies, consulates

Operators

Visa operators, airlines, partners

Documents

PDFs, rules, notices and forms

VFS Travel Intelligence Engine

Discovery loop

Find, fetch, validate and qualify new sources, then add them to the source registry.

Verified knowledge base

Extracted, harmonised and grounded facts, stored as a graph with source, authority and fetch time.

Monitoring loop

Re-check approved sources, detect material change, re-extract and send it for human review.

Used by

Travellers

Clear answers, with sources

VFS teams

Across internal channels

Partners

Through an API

VFS owns the system, the knowledge base and every review decision. It becomes a proprietary asset that improves with each update.

07 / 16
03The system

Finding and reaching the right sources

Discovery never stops: each run finds new sources and improves the registry. The crawler always tries the cheapest way in first.
Discovery: from a question to a trusted source

Find

Search the VFS estate, mission sites, visa operators and immigration sites.

Reach

Fetch and inspect HTML pages, APIs, PDFs and dynamic pages.

Validate

Right page, relevant, accessible, not a duplicate, not an error page.

Qualify

Check authority, relevance to the corridor, coverage and purpose.

Register

Store URL, authority tier, purpose, fetch method and last check.

Access: escalate only when a source requires it

Direct HTTP

Static pages and APIs. Fast, lightweight, low cost.

Browser rendering

JavaScript-heavy pages, rendered as a user sees them.

Special access

Blocked or location-sensitive sources: stealth browser, residential or in-country access.

Every fetch is validated

A successful response is not enough. We confirm the content is correct, relevant, not blocked and not an empty placeholder.

08 / 16
03The system

From raw pages to verified facts

Rules handle what is structured, models handle the rest, and no claim ships without evidence in the source.
Step 1

Parse and harmonise

Structured content is mapped with deterministic rules. Only unstructured content goes to a model, and its output is validated against a fixed schema.

Only the sources a question needs are processed, and content is compressed first. One extraction serves business, tourism and transit alike.

Step 2

Ground every claim

An independent verifier checks each claim against the text the system actually fetched.

Seafarer Visa is required to enter.Keep
Matching text found on vfs.global

Digital Nomad Visa is available.Drop
No supporting text in any source

Step 3

Store as a graph

Countries, visa types, requirements, validity and sources are linked, not filed as separate documents.

Every fact carries its authority: mission and government for policy, VFS and operators for application steps, reference sources only where the primary source is silent.

The rule
If a claim cannot be found in the source evidence, it does not ship.
09 / 16
03The system

Keeping every answer current

Once a source is approved, the system acts only when it changes. Effort grows with the amount of change, not with the number of pages.
Monitoring: act only on change

Check

Daily or weekly per source, using conditional requests so unchanged pages are not downloaded again.

Detect

Compare content hash, timestamp and meaning against the last version.

No change

Keep the existing record. Nothing else runs.

Cosmetic

Log it. No re-extraction, no review.

Material

Re-extract, send for human review, then update the record.

Freshness: every answer shows its age

Defined

Each source type has a freshness target: daily for government and operator pages, longer for stable sources.

Traceable

Every fact carries its fetch time.
VFS Global Switzerland · checked 16 Sep 2026, 03:04 IST

Auditable

Material changes are stored with the difference.
“until 30 June” → “until 31 July”
Recorded 11:20 · reviewed 13:45 · published 14:10

10 / 16
03The system

Cost stays low by design

Process less, use the smallest model that does the job reliably, and serve answers without calling a model at all.
Measured
99.93%

less data to process

8.97M tokens of raw source pages reduced to 6,718 tokens of relevant, structured content before any model sees it.

0

model calls when answering

$550 to $1,450

estimated running cost a month, depending on re-check frequency

Cheapest method first

Deterministic rules

The default: validation, compression and serving, wherever the problem is well-defined.

Small models

SLMs and VLMs for bounded extraction and understanding tasks.

Domain models

Regulatory, legal and multilingual content.

Frontier models

Only for ambiguous cases that need deeper reasoning.

Cost figures are estimates from current benchmarks and will be refined as measured rates mature.

11 / 16
04Economics

Economics at scale

Running cost follows the sources that change, not the number of questions asked. Covering VFS’s own estate costs about $90 a year.
Illustrative annual running cost
ScenarioCorridorsSource mixPer year (USD)Notes
VFS estate only1,826100% VFS~$90VFS content API. No model calls.
Priority expansion5,00030% VFS, 70% external~$3,900High-value countries and services.
Broader coverage10,00020% VFS, 80% external~$7,900More countries and source types.
Global scale39,8005% VFS, 95% external~$32,300The full corridor universe.

Build once

Source discovery, initial fetch and extraction, knowledge base and pipelines, environment, security and monitoring. Optional private GPU cluster: about US$11,000 for 2 nodes.

Run continuously

Monitoring and re-fetching, model use for changed content only, egress where needed, storage and serving, human review. Indicative: US$550 to 1,450 a month.

Illustrative estimates from current benchmarks; actual cost depends on source mix, content complexity and usage. Engineering is priced separately in the engagement options.

12 / 16
04Economics

VFS owns the intelligence

Models are interchangeable. VFS’s knowledge stays in VFS’s environment whichever model runs underneath, with no dependency on a single provider.
VFS owns
PromptsExtracted corpusKnowledge graphSource registryReview decisions

Served by VFS

The knowledge base is served from VFS infrastructure, for example Cloudflare.

No model at request time

Answering is a database read, not model inference.

Data residency

Can be hosted in-region where required.

Infrastructure path: move only when volume justifies it
Today

Together AI

Open-weight models, serverless, already used by VFS. No commitment.

Next

AWS Bedrock, Mumbai

Inside VFS’s AWS environment, when volume or data residency calls for it.

At high scale

Dedicated GPU

Only at sustained high use, around 5M tokens an hour.

If required

Own hardware

Possible, but not the default for a bursty workload like this.

13 / 16
05Delivery

From prototype to the VFS estate

Start with the 1,826 corridors VFS already publishes, ingested directly through the VFS content API, then expand to external sources.
Phase 1 · Done
31

Prototype validated

End-to-end pipeline proven across diverse countries and source types.

Phase 2
1,826

VFS estate integrated

235 origins and 1,446 visa corridors, through the content API.

Phase 3
5,000+

Priority expansion

High-value countries and services from external sources.

Phase 4
39,800

Global coverage

The full corridor universe, over time.

VFS estate: direct and cheap

The content API gives structured access to VFS-published content. No browser rendering, no residential egress, no model calls for retrieval or change checks.

Everything else: discovery and crawler

Government, mission and other operator sources. Both paths feed the same harmonisation, grounding, monitoring and knowledge base.

14 / 16
05Delivery

Who is building it: 360 Labs

An AI engineering and research lab. We build production systems where information is fragmented, keeps changing, and has to drive an operational decision.

Retail F&B

One of India’s largest retail F&B companies

Location-aware crawler. Mapped every customer pin code to the right dark store, selected inventory that could actually serve it, and validated prices against Instamart.

Location → source → availability → validation

DSV

World’s largest logistics company

Fineasy, billing and documents. Extracts TMS and POD details with OCR, validates them, calculates GST, routes approvals and generates self-invoicing PDFs. SAP integration is the latest extension.

Documents → extraction → validation → workflow

Pine Labs

Merchant commerce and fintech platform

Merchant risk and retention. Combines signals across 2M+ merchant touchpoints to spot disengagement early, explain what changed and trigger targeted interventions.

Signals → change detection → explanation → action

Our own research

Research that feeds our engineering

SLM360 small models that run on devices. Aura a programming language and compiler. Omni systems for data understanding and forecasting.

15 / 16
05Delivery

Engagement and next steps

Two options. The same 2-month build by a 4-person team; the difference is who runs the system afterwards.
Option 1 · Build and handoverOption 2 · Build and maintain
Build$26,000$26,000
Team and timelineAI-native engineering lead, AI-native PM and 2 full-stack AI engineers. 2 months.Same team and timeline.
After the buildStructured handover and knowledge transfer to your team.$2,500 a month for 2 full-stack AI engineers: maintenance, updates and continuous improvement.
Next steps · we can start immediately

1  Scoping call

Confirm priorities, success criteria and requirements.

2  Access

Key stakeholders, systems and documentation.

3  Kick-off

Start the delivery pod and detailed design.

4  Deliver

Build in 2 months, then handover or maintenance.

16 / 16
Live prototype
chancery.360labs.tech/sandbox
New tab
All slides
Arrow keys to navigate  ·  F present  ·  G all slides