How the platform is put together, and why it is put together that way.
Four systems on one platform, built on the same spine. What changes between them is the domain and the agents. What does not change is where the boundaries sit.
Entry
What starts a run
Campaign brief
Traveller chat
Brand site
Agents
Specialists that each own one craft
Director
Script writer
Photographer
Cinematographer
Profile
Planner
Booking
LangGraph · LangChain · MCP · Function calling
Domain
Business logic, independent of any provider
Campaign
Itinerary
Brand surface
Python · FastAPI · Node.js
Adapters
Swappable, held at the edge
LLM providers
Media generation
Booking APIs
Structured data
OpenAI · Anthropic · Gemini · Chroma · pgvector
Infrastructure
Terraform, provisioned against demand
Environments
Instance lifecycle
Deploy
Terraform · AWS · GCP · Docker · GitHub Actions
The layering is deliberate. Agents sit above the domain and never below it, so an agent can be rewritten without the business logic noticing. Providers sit at the edge behind adapters, so changing a model is a configuration change, not a migration. Infrastructure is described in the same repository as the thing it runs.
That is what let four products share one platform instead of becoming four codebases. The media house and the destination desk have almost nothing in common at the domain level, and almost everything in common below it.
PRINCIPLES BEHIND IT
Design for the failure modes
The interesting problems are rarely the model call. They are retries, partial results, cost, and knowing when an agent should stop rather than keep trying.
Boundaries are the deliverable
Deciding what belongs inside the domain and what stays at the edge is the actual work, and holding that line when a deadline argues otherwise is most of it.
Prototype and production share a codebase
A demo that has to be rewritten to ship was never evidence it would ship. The path to production is a design constraint, not a later phase.
Write for machine readers too
Increasingly the first reader of anything you publish is a model. How a system describes itself now shapes how it is found and repeated.
02
Systems
What that architecture is carrying.
Automated media
In production
An agent-orchestrated media house
A director, a script writer, a photographer and a cinematographer, running as agents. Together they produce publishable marketing media end to end.
StepAgent
Modelling agents on production roles is what makes it work. A director owns coherence and delegates to specialists that each own one craft. That holds up far better than one general-purpose agent, and every role can be improved on its own.
Generation providers stay at the edge of the system. Models change constantly, so the domain logic should not change with them.
AI AgentsAgentic WorkflowsLLM SystemsAutomated MediaDomain-Driven Design
Production
Running for wine brands
Text · Image · Video
Generated modalities
Destination management
In production
An automated destination management desk
The work a destination management agent does by hand. Profiling a traveller, planning the trip, sourcing and booking each activity, running as a chain of agents.
A profile agent reads the conversation for what the traveller actually wants. A planner agent turns that into a day by day itinerary. A booking agent routes each activity to the relevant provider, asks whether it can be had, and books it.
Every activity carries its own state. It starts pending, and only a successful booking moves it to confirmed. Otherwise it is cancelled, so the itinerary describes what is actually held rather than what was proposed.
AI AgentsConversational AIStructured OutputState MachinesThird-party APIs
Infrastructure as code
Demand-driven instance automation
Terraform provisions and de-provisions instances for digital brand solutions, scaling with the demand of each brand workflow.
Instances come up and go down as workload requires, with lifecycle handling robust enough to run unattended. Provisioning expressed as code lets capacity track demand without becoming fragile.
TerraformInfrastructure as CodeCloud ArchitectureAutomation
Discoverability
Current focus
AI SEO for brands
Making a brand legible to the systems that now answer questions about it. Structuring a site so an answer engine retrieves and cites it correctly, not just so a crawler ranks it.
A crawler and an answer engine read the same site for different reasons. One builds an index to rank. The other retrieves passages to summarise and attribute. Optimising for the first is twenty years old. Most brands are invisible to the second.
The work is to give a site a machine-readable surface. Schema.org data describing what the brand actually is, semantic markup that survives extraction, explicit machine-readable context. The real question becomes what a model says about the brand when nobody is checking, which is closer to brand safety than marketing.
AI SEOAnswer Engine OptimisationStructured DataSemantic MarkupSchema.org
Building the AI systems behind WineConX’s digital marketing platform for the wine industry: an agent-orchestrated media production pipeline, a conversational itinerary system, the Terraform-managed infrastructure they run on, and AI SEO for client brands.
Own the AI systems end to end: agent design, the architecture around them, and the infrastructure they run on.
Applied domain-driven design so model providers can be swapped without reworking the core.
Took systems from prototype to production, on Terraform environments that run unattended.
AI AgentsAgentic WorkflowsLLM SystemsSoftware ArchitectureDomain-Driven DesignTerraformAI SEONode.js
Mar 2024 to Feb 2026
Research Assistant
FZI Research Centre · Karlsruhe, Germany
Applied research on multimodal LLM systems and graph neural networks, spanning automated UX evaluation and epidemiological risk prediction.
Nov 2024 to May 2025
Master's Thesis: EffiComm
FZI Research Centre · Karlsruhe, Germany
Bandwidth-efficient multi-agent communication for cooperative autonomous driving, published at IEEE ITSC 2025.
Feb 2023 to Mar 2024
Software Developer (Werkstudent)
Vanory · Karlsruhe, Germany
RTOS firmware and edge AI for consumer IoT lighting hardware shipping at production volume.
50x
Power reduction, 1A to 20mA
500+
Devices in production
Jan 2023 to Jan 2024
Software Developer (HiWi)
RITA Project · Karlsruhe, Germany
Computer vision and photogrammetry for terrain understanding in medium-scale robotic systems.
EARLIER
Mar 2022 to Sep 2022Associate Software DeveloperGupshup · Mumbai, India
Jan 2022 to Feb 2022Machine Learning InternLife Spark Technology, IIT Bombay · Mumbai, India
Aug 2021 to Dec 2021Embedded Systems DeveloperMetwiz Materials · Maharashtra, India
04
Selected work
A few projects that show the range, from cooperative perception to microcontroller-scale inference.
IEEE International Conference on Intelligent Transportation Systems (ITSC) · 2025
Graph neural networks with a mixture-of-experts architecture cut vehicle-to-vehicle perception bandwidth by 83.7% while holding 92% detection accuracy.