Your data stays yours.
Design ingestion, access boundaries, and retrieval around your requirements. Keep sensitive information in an environment you control.
Your intelligence.
Your infrastructure.
Your AI.
AI that belongs inside your business. I help teams build and fine-tune models on their own data, deploy in their own environment, and take control of the whole system.
BLUEPRINT / 001● SOVEREIGN BY DESIGN
▤
Your training dataCompany-owned datasets · domain expertise
⌘
Your model weightsCustom fine-tuning · evaluated on your tasks
▣
Your environmentPrivate cloud · on-premises · your controls
BOUNDARY: YOUR ORGANIZATION↳ BUILT TO HAND OVER
An AI system you can understand, operate, and keep.
95%support automation on an omni-channel agentic CX system
75%lower inference latency, with 95% accuracy maintained
15%accuracy gain from fine-tuning ML models and LLMs/VLMs
Gold MedalChairman's Gold Medal, LNMIIT, for performance across all facets of college life
Flagship offering / Sovereign AI
Own the system.
Keep the know-how.
Your competitive knowledge should become a capability your team can keep building on. Ownership starts with architecture and ends with a usable handover.
Models that know your domain.
Prepare datasets, adapt suitable open-weight models, and evaluate them against real tasks. Use fine-tuning where it helps and retrieval where knowledge needs to stay current.
A complete handover.
Source code, model artifacts, evaluation suites, deployment configuration, and documentation, so your team can operate and evolve the system.
Selected work
Systems I've built
and what they changed.
Production AI from my roles at Chatarmin, Curiflow, and DataCraftAI. Client details stay confidential.
Agentic customer support across WhatsApp, Instagram, and web chat
An autonomous, omni-channel CX system for enterprise e-commerce clients. I designed the multi-agent orchestration layer for intent routing, state transitions, context sharing, tool execution, and handoffs between specialised agents, external LLM providers, and human support teams. I also built in-house memory combining session state, long-term customer knowledge, and RAG/Graph-RAG retrieval, plus evaluation pipelines for retrieval accuracy, response quality, latency, and agent failures.
95%support automation
Document understanding and domain-adapted models
End-to-end document understanding pipelines with automated evaluation, in compliance-focused environments. I trained and fine-tuned traditional ML models and LLMs/VLMs, built chat-based query engines on self-hosted open-source LLMs with tool access to SQL/NoSQL databases, and deployed on AWS/GCP with Docker and Kubernetes.
15%accuracy improvement from fine-tuning
Faster inference for document analysis
Worked on document analysis and query-understanding pipelines, and optimized their inference path. Experimented with NLP and computer vision models to speed up iteration and prepare them for deployment.
75%lower latency at 95% accuracy
Conversational automation for real estate agents
Integrated LLMs with WhatsApp APIs and webhooks to automate agent communications and transaction workflows. Designed the product architecture and backend systems for scalable AI integration.
Less manual workin agent communications
BEHIND THE WORKBENCH / VK
vk.
MUMBAI, INDIA · REMOTE
Hi, I'm Vaibhav.
I like connecting the model
to the real-world problem.
I'm a Machine Learning Engineer focused on LLMs, multi-agent orchestration, NLP, retrieval, and conversational AI. I build customer-facing systems that have to work in production, in privacy- and compliance-sensitive environments.
Today I'm a Research Engineer at Chatarmin, a Vienna-based company, building agentic customer support for enterprise e-commerce. Before that, at Curiflow in Mumbai, I built document-understanding pipelines, fine-tuned models, and self-hosted LLM query engines.
I studied Computer Science at LNMIIT, Jaipur, where I received the Chairman's Gold Medal. I was also named Best All-Rounder by Bank of Baroda, and served as President of the Student's Gymkhana, representing 2000+ students.
PYTHON · C/C++ · SQL · TYPESCRIPT
PYTORCH · CUDA · XGBOOST
RAG & GRAPH-RAG · AGENTS
AWS · GCP · DOCKER · KUBERNETES
Services
What I build.
Alongside Sovereign AI, I take on focused projects in each of these areas.
Custom models & machine learning
Dataset preparation, model fine-tuning, computer vision, and NLP. From classical ML to transformer-based systems, selected around the problem.
OCR & document intelligence
Turn difficult documents into usable information. Handle layout, tables, visual questions, and entities with evaluation at each stage.
RAG, memory & ingestion
Connect your organization's knowledge to its AI systems. Build ingestion pipelines, retrieval, and memory for relevant, grounded context.
Multi-agent systems & chatbots
Build purpose-specific assistants and coordinated agents with tool access, shared context, state management, and human handoffs.
Evaluation, optimization & deployment
Measure quality on your tasks, investigate failure modes, optimize inference, and deploy with observability. Make performance and operating costs visible.
Explore the economics
Make the numbers work.
Three ways to assess an AI investment. Each starts with an illustrative example; replace it with your own measurements.
Methodology & sources
Own AI: upfront = data + engineering + evaluation + training compute. Training compute = GPUs × hourly GPU rate × hours per run × runs. Monthly hosting = replicas × billed hours × hourly replica rate, plus storage, operations, and retraining. Alternative cost = fixed monthly cost + successful tasks × per-task price. Compare the same quality, workload, and service requirements.
Automation: adopted tasks = volume × eligible share × adoption. Successful attempts retain review and rework time. Failed attempts retain original handling time plus exception overhead. Unadopted tasks retain original handling time. Time savings: volume × adoption × (minutes before − minutes after) ÷ 60.
Capacity value = released hours × loaded hourly cost. Return on capacity treats all released time as value, which is useful when the time is redeployed to other work. Cash savings count only the spending-reduction percentage you enter. Added labour is counted in full. ROI = (benefits over period − upfront − recurring costs over period) ÷ (upfront + recurring costs over period). Simple payback = upfront ÷ positive monthly net benefit, assuming steady cash flows.
These are planning estimates, not quotes or forecasts. Example amounts are not converted between currencies.
Sources: AWS: generative AI costs · NVIDIA: inference benchmarking · Microsoft: measuring business value.
Private AI & governance
Control your AI.
Design for your obligations.
Private deployment gives you choices over data access, location, and retention. I scope access controls, encryption, data minimization, retention and deletion, audit logging, human oversight, and evaluation evidence alongside your legal and security teams.
Deployments can be designed around frameworks such as
- EU GDPR
- UK GDPR
- CCPA / CPRA
- India DPDP Act
- HIPAA
- PCI DSS
- EU AI Act
- ISO/IEC 42001
- NIST AI RMF
- ISO/IEC 27001
- SOC 2
Engineering support for your compliance work, not legal advice or a certification. Base models and third-party software keep their original licenses; deliverables are agreed per engagement.
Working together
A clear path to ownership.
Pick the workflow.
Find the manual work worth automating and measure its time, cost, and quality baseline.
Test the approach.
Build a focused prototype and evaluate it on representative tasks.
Make it operational.
Integrate, optimize, and deploy into the agreed environment.
Make it yours.
Deliver the agreed artifacts, documentation, and operational walkthrough.
The first step
Start with one workflow.
We choose a single workflow, measure its baseline, and build a focused pilot. You decide on a full build once the pilot's results are in.
Your next capability starts with a conversation
Let's build AI you can own.
Tell me what your team does, where the friction is,
and what needs to stay under your control.