VFS Travel
Intelligence Engine
VFS Global
Enabling secure and seamless mobility for a better tomorrow
360 Labs
Applied AI for a more open world
September 2026
Private and confidential
Map every corridor, maintain the intelligence
potential corridors
200 countries × 199 possible destinations. A global, dynamic problem across almost every origin and destination combination.
Information is fragmented and changes over time
VFS, governments, missions, operators, PDFs and operational notices all contribute pieces of the answer.
Map the corridor universe
Discover relevant sources, resolve who actually handles each corridor, and build a structured corridor record.
Maintain the information
Monitor approved sources, detect changes, verify evidence and update the trusted record.
corridors
29 Schengen + US + NZ
grounded claims
510 / 512 supported by source evidence
visa types
Purpose-tagged across the corpus
model calls at request time
Served from pre-built, verified intelligence
Built and measured against live VFS and external government and operator sources.
AI Engineering & Research Lab
| Where AI sits | Product | What it does |
|---|---|---|
| AI inside workflow | Travel OS | Classifies enquiries, drafts responses and routes cases. |
| AI inside OS | Manufacturing CRM + ERP | Predicts, recommends and automates work across core operations. |
| AI inside the intelligence layer | Business Brain | Turns operational data into patterns, findings and decisions. |
| Research | What it enables |
|---|---|
| SLM360 Smaller, efficient language models | AI that can run directly on devices. |
| Aura New programming language and compiler | Faster, more controlled software development. |
| Omni 1.0 / OmniScient AI systems for data understanding and forecasting | Better answers and forecasts from business data. |
Research
Find what’s missing
Engineering
Build the system
Production
Make it reliable
Deployment
Run it at scale
Relevant Work
Retail F&B
One of India’s largest retail F&B companies
Location-aware web crawler
Mapped each customer pin code to the right dark store, selected the inventory that could actually serve that location, and validated prices against Instamart.
DSV
World’s largest logistics company
Fineasy: billing & document workflow
Turned shipment documents into validated invoices: extracting TMS and POD details with OCR, checking them, calculating GST, routing approvals and generating self-invoicing PDFs. SAP integration is the latest extension.
Pine Labs
Leading merchant commerce & fintech platform
Merchant risk & retention intelligence
Combined signals across 2M+ merchant touchpoints to identify early signs of disengagement, explain what was changing and surface targeted interventions.
Government · Logistics · Manufacturing · Defence · Real Estate · Construction · Education · Healthcare · Fintech · Maritime · Consumer · Retail · E-commerce · Aviation & Aerospace · Energy · Legal · Robotics
One travel question can depend on information spread across multiple systems, owners and documents.
VFS website
Country pages, one-pagers
Mission / Government sites
Notices, policy updates
Other operators
Some corridors not operated by VFS
Supporting documents
Checklists, PDFs, application rules
Different answers to different parts of the same question.
Interpretation and validation stays with people.
Source fragmentation
The answer is rarely on one page. 31 corridors examined.
Ownership & jurisdiction
The organisation serving the corridor is not always obvious. 4 of 31 operated by someone other than VFS.
Semantic variation
The same concept is described differently across sources. 212 distinct visa type names.
Continuous change
The source of truth can change without the answer changing with it. 95 operational notices found.
The challenge is not collecting information.
possible origin to destination corridors
pages to process
Planning assumption: 10 pages per corridor
diplomatic posts
In the upper-bound source universe
Discover
Find the right sources
Access
Reach different source types
Understand
Turn different information into one record
Verify
Prove every claim
Monitor
Detect meaningful changes
Review
Put people where judgement matters
A working prototype tested the core intelligence pipeline on real VFS, government and operator sources across 31 corridors.
corridors
29 Schengen + US + New Zealand
sources across VFS, governments and operators
PDFs: application rules, checklists, notices and forms
- VFS websites Country pages, one-pagers
- Government sources Embassy and consulate sites
- Other operators Third-party visa operators
Discover
Find and validate relevant sources
Harmonise
Turn different source structures into one usable record
Ground
Tie each claim to source evidence
Materialise
Store the approved result for serving
A single, structured knowledge base
Current, source-linked and ready to serve
corridors tested
claims grounded in source evidence
visa types purpose-tagged
model calls at request time. Answers are served from the knowledge base, not an LLM.
The working prototype
Live prototype
The sandbox could not be loaded here. It may be offline, or you may have no network connection.
Open chancery.360labs.tech/sandboxThis is the live sandbox. Type a question and use it directly. Use the arrow keys or the bar below to move between slides.
How the system works
- VFS websites
Country pages, visa information - Government sources
Ministries, embassies, consulates - Operator sources
Airlines, travel operators, partners - Documents
PDFs, application rules, notices and forms
Discovery, intelligence and monitoring work together to build a trusted, up-to-date knowledge base.
Discovery
Find and qualify new sources
- Identify potential sources
- Fetch and validate content
- Assess and qualify
- Add to source registry
Travel intelligence
Structured, source-backed, human-approved
Monitoring
Keep approved information current
- Check approved sources
- Detect meaningful changes
- Assess impact
- Re-extract when required
- Human review before publishing
- Travellers
Get clear, reliable answers - VFS teams
Use across internal channels - Partners
Integrate via API
VFS owns the system, the structured knowledge, processing logic and accumulated review decisions.
A high-quality, structured travel intelligence asset that compounds over time.
New sources and changes make the knowledge base richer and more valuable.
Every answer is backed by source evidence and human review.
Discovery
Find
Search for potential sources
- VFS estate
- Government / mission websites
- Visa operators
- Policy and immigration sites
- Other relevant sources
Reach
Fetch and inspect what was found
- HTML pages
- APIs
- PDF documents
- Dynamic pages
- Other formats
Validate
Check whether the source is real and usable
- Correct page
- Relevant content
- Accessible
- Not a duplicate
- Not a placeholder or error page
Qualify
Decide whether the source should be trusted
- Authority and ownership
- Relevance to the corridor
- Coverage of key topics
- Purpose (application, rules, schedules, etc.)
Register
Add approved source to the source registry
- Source URL
- Authority tier
- Purpose / category
- Fetch method
- Last checked
A trusted and traceable library of approved sources.
| Source | Authority | Purpose |
|---|---|---|
| India Mission (Germany) | Government | Jurisdiction |
| VFS Global | Operator | Application |
| TLScontact | Operator | Intake |
| Germany Immigration | Government | Visa rules |
| … | … | … |
Every run uncovers new sources and improves the quality and coverage of the source registry.
The system keeps looking for new sources, even after the initial set is built.
Monitoring
Approved source
From the source registry built in Loop 1.
Check
On a defined cadence: daily or weekly. Conditional requests, adaptive frequency.
Detect change
Compare with the last version: content hash, updated timestamp, semantic assessment.
Keep existing intelligence. No further action.
Record the change. No re-extraction. No human review.
Re-extract latest content → human review → update intelligence.
Full pages are not re-processed when nothing has changed.
The system first checks whether the content has changed. Where supported, conditional requests avoid downloading the page again.
Expensive processing is reserved for meaningful changes.
A changed page is assessed before re-extraction. Cosmetic changes stop without consuming the heavier processing path.
Human review is reserved for changes that require judgement.
Only material changes enter the review queue. Everything else is handled automatically.
The Crawler
Direct HTTP
For static pages and APIs.
- HTTP → HTML / JSON
- Fast, lightweight and low cost
Browser rendering
For JavaScript-heavy pages.
- Renders the page as a user would see it
- Handles dynamic content
Special access
For blocked or location-sensitive sources.
- Stealth browser / residential or in-country access
- Used only when normal access is not sufficient
A successful response is not enough. We check:
Correct content
Not an error page
Relevant page
Matches the source and topic
Not blocked
Detects access restrictions
Not a placeholder
Avoids generic or empty shells
Parsing & Harmonisation
Information comes in many formats, structures and levels of clarity.
- Operator sites
Application steps, requirements and service information. - Government sites
Visa and jurisdiction policies. - Documents
Checklists, notices and supporting files.
AI when needed.
The model is not the parser for everything.
Structured information
Use deterministic rules. Map, standardise and store directly.
Unstructured information
Extract with a model, then validate. Extract only what’s needed and validate against a defined schema.
Selective processing
Only the sources needed for the question are sent to the model. Keeps the output accurate and efficient.
Travel intelligence record, stored in a standard format.
| Visa type | Short stay |
| Purpose | Tourism |
| Application | VFS / operator |
| Jurisdiction | United Kingdom |
| Key requirements | Valid passport Proof of funds Return ticket |
| Source quote | “You must show proof of sufficient funds …” vfs.global/st-visa |
One extraction serves multiple purposes
Business, tourism and transit are filters on the same underlying catalogue, rather than separate extractions.
Only relevant information reaches the model
Different question types use different source sets, keeping the processing focused and accurate.
Compression happens before model processing
Content is filtered and reduced to what’s needed, keeping costs under control while retaining the required facts.
Grounding
Each claim is checked against the actual content of the sources.
Seafarer Visa
“Seafarers must hold a valid Seafarer Visa. Book an appointment and submit the required documents…”
Example source fetched by the system. Government, VFS or operator page.
Seafarer Visa
Seafarers require a visa to enter.
Exact or closely matching content found in source.
Digital Nomad Visa
A digital nomad visa is available for this country.
No supporting content found in any source.
of claims grounded
claims supported by source evidence
unsupported claims were rejected
Evidence is stored with each claim, including where it came from and what authority it carries.
Mission / Government
Jurisdiction and policy.
VFS / Operator
Application and service information.
Reference sources
Used where the primary source does not state the required fact, with provenance clearly attributed.
Knowledge Graph
Each verified claim is structured and stored with its source and authority.
- VFS / Operator
Application steps, document requirements, service details - Mission / Government
Visa policy, jurisdiction rules - Reference sources
Used where primary source does not state the required fact
and linked
Information is stored as a graph, linking countries, visa types, requirements and sources.
and consistent answers
The same connected data serves different questions and user needs.
- Answer user queries
Get accurate, sourced answers. - Compare options
e.g. visa types, requirements across countries. - Track changes
See what has changed and when. - Enable analytics
Identify trends and gaps.
Source-aware
Every fact is stored with its source and authority.
Connected, not isolated
Relationships between countries, visa types, requirements and sources are linked.
Reusable
The same verified data supports multiple journeys and product features.
Freshness
Defined
A freshness target is set for each source type.
Government & operator pages
Checked regularly to capture changes as they happen.
Less volatile sources
Checked at a longer interval where appropriate.
Defined freshness expectations
Know how often important sources are checked.
Traceable
Every piece of evidence carries its fetch time.
VFS Global, Switzerland
“Short-stay visa applications are accepted at the VFS Global centre in New Delhi, Mumbai, Chennai, Kolkata and Bengaluru.”
16 Sep 2026 · 03:04 IST
Timestamped evidence
Know when the information behind an answer was last verified.
Auditable
Material changes are recorded with the difference.
Applications accepted until 30 June.
Applications accepted until 31 July.
| Change recorded | 02 Jul 2026 · 11:20 IST |
| Reviewed | 02 Jul 2026 · 13:45 IST |
| Intelligence updated | 02 Jul 2026 · 14:10 IST |
A record of material change
Know what changed, when it changed and what was approved.
How We Control Cost at Scale
We process less, use models selectively, and serve without a model.
tokens · raw content from source pages
tokens · structured, relevant information
less data to process
(measured benchmark)
Answers are already stored and served from the knowledge layer.
Infrastructure + model costs, depending on how aggressively sources are re-checked.
Figures are estimates based on current assumptions and will be refined as measured rates mature.
The right approach for the right task. Cheapest first.
Rules, validation, compression, serving.
Used wherever the problem is well-defined.
Used for suitable, bounded tasks.
Helpful for specific extraction and understanding tasks where appropriate.
Used where the problem justifies a specialised model.
For complex, domain-heavy content (e.g. regulatory, legal, multilingual).
Used for harder interpretation when needed.
For ambiguous or highly complex cases that require deeper reasoning.
1. Fetch
Only when required
2. Process
Only relevant information
3. Model
Only where needed
4. Review
Only material changes
5. Serve
Database read (no model call)
Economics at Scale
Build once
CapexOne-time investment to set up the platform, integrate sources and establish the infrastructure.
| Source discovery and qualification | Ingest and structure initial set of sources |
| Initial fetch and extraction | Build the knowledge base and pipelines |
| Infrastructure setup | Environment setup, security, monitoring |
| Private GPU option (optional) | ~US$11,000 for 2-node cluster Available through 360 Labs or distribution partner. One-time infrastructure acquisition. |
Engineering and implementation (e.g. 2-month build) are priced separately (see Engagement and Next Steps).
Run continuously
OpexOngoing costs to monitor sources, process changes and keep the intelligence layer up to date.
| Source monitoring and re-fetching | Scheduled and change-based checks |
| Model usage for changed content | Only for new or updated content |
| Browser / residential egress | Where required for external sources |
| Storage and serving | Database, vector store and application infra |
| Human review | Review of material changes with audit trail |
| Indicative monthly cost | ~US$550 to 1,450 / month Depends on monitoring strategy, source-change behaviour and infrastructure choice. |
Based on phased coverage across the 39,800 corridor universe.
| Scenario | Corridors covered | Mix assumption | Est. annual cost (USD) | Notes |
|---|---|---|---|---|
| VFS estate only | 1,826 | 100% VFS estate | ~$90 | Uses VFS content API. No model calls. |
| Priority expansion | 5,000 | 30% VFS / 70% external | ~$3,900 | Focus on high-value countries and services. |
| Broader coverage | 10,000 | 20% VFS / 80% external | ~$7,900 | Expands to more countries and source types. |
| Global scale | 39,800 | 5% VFS / 95% external | ~$32,300 | Full corridor universe, mixed source types. |
Infrastructure Strategy
Prototype to production
Together AI
Open-weight models. Serverless. Already used by VFS.
When volume or residency justifies it
AWS Bedrock (Mumbai)
Inside VFS’s AWS environment. Best fit when volume or data residency justifies it.
Only when utilisation is consistently high
Dedicated GPU
Makes sense only at sustained high utilisation (~5M tokens/hour).
Not the default
Own hardware
Possible, but not the default for this workload due to bursty usage and operational constraints.
Ownership
The intelligence layer built for VFS
VFS-specific assetsBest available models for the workload
Provider infrastructureAnswers are served from VFS’s own environment
Under VFS controlMaterialised knowledge base
Served via VFS infrastructure (e.g. Cloudflare).
0 model calls at request time
Database read, not model inference.
Data residency
Can be hosted in-region where required.
From Prototype to VFS Estate
corridors tested. Prototype validated.
End-to-end pipeline validated across diverse countries and source types.
distinct corridors
origins
visa corridors
Excludes attestation and permit services
Represents 4.6% of the ~39,800 corridor universe.
VFS content API
Access VFS-published content directly via API. No browser rendering. No residential egress. No model calls for retrieval or change checks.
External source discovery and crawler
For government, mission and other operator sources.
Both paths, one system
Both paths feed the same harmonisation, grounding, monitoring and knowledge system.
31 corridors
Prototype validated
VFS estate integrated
1,826 corridors via API
Priority expansion
High-value countries and services
Broader global coverage
Full 39,800 corridor universe over time
Engagement and Next Steps
Option 1 Build + HandoverA focused build to deliver a production-ready system, followed by a structured handover to your team. |
Option 2 Build + MaintainWe build the solution and continue to support it with a lean engineering team for ongoing maintenance and iteration. | |
|---|---|---|
| Build price | $26,000 | $26,000 |
| Delivery team | 4 members AI-native Eng Lead, AI-native PM, 2 full-stack AI engineers | 4 members (build) AI-native Eng Lead, AI-native PM, 2 full-stack AI engineers |
| Delivery timeline | 2 months | 2 months |
| Maintenance | Not included | $2,500 per month 2 full-stack AI engineers (post-build) |
| Outcome | Fully built solution with knowledge transfer and handover. | Production-ready solution with ongoing engineering support for stability, updates and continuous improvement. |
We can move forward immediately. The first step is a short scoping call to align on priorities and access.
Align on scope
Confirm priorities, success criteria and any specific requirements.
Share access
Provide access to key stakeholders, systems and documentation.
Kick-off
Start the delivery pod and begin detailed design.
Deliver
Complete build in 2 months, followed by handover (Option 1) or transition to maintenance (Option 2).