# Marquedata

> Marquedata builds modern data platforms, AI and GenAI solutions, customer intelligence systems and intelligent software products that help businesses make faster, smarter decisions.

Marquedata is an independent data and AI engineering company based in Kolkata, India, working with organisations in India and worldwide. Positioning: "Building Intelligent Data & AI Solutions for the Modern Enterprise". Contact: info@marquedata.com.

Marquedata (marquedata.com) provides data engineering, cloud data platform, data science and machine learning (ML model training, LLM fine-tuning, MLOps), generative AI and agentic AI, analytics and decision intelligence, and AdTech / MarTech customer-intelligence services, and is developing its own B2B SaaS products. It is not affiliated with data-storage hardware products or other similarly named businesses.

## Solutions
- [Data & Cloud](https://marquedata.com/solutions/data-cloud/): Modern, governed data platforms on the cloud — engineered for scale, reliability and AI readiness. Capabilities: Data Engineering, Data Architecture, Cloud Data Platforms, Databricks, Apache Spark, Kafka & Streaming, ETL / ELT, Data Warehousing, Data Quality, Data Governance, Data Observability, Data Lineage.
- [AI & GenAI](https://marquedata.com/solutions/ai-genai/): Generative AI, RAG and agentic systems engineered for enterprise reliability, security and control. Capabilities: Generative AI, Retrieval-Augmented Generation, Vector Search, Knowledge Graphs, Enterprise Search, AI Agents, Agentic AI, Human-in-the-loop, AI Automation, Machine Learning, AI-powered Analytics, Evaluation & Guardrails.
- [Data Science & ML](https://marquedata.com/solutions/data-science-ml/): Data science, machine learning, model training and LLM fine-tuning — from experiment to monitored production model. Capabilities: Data Science, Machine Learning, Feature Engineering, ML Model Training, Deep Learning, LLM Fine-Tuning, Training Data Curation, Model Evaluation, MLOps, Model Deployment & Serving, Model Monitoring, Responsible ML.
- [Analytics & Intelligence](https://marquedata.com/solutions/analytics-intelligence/): From business intelligence to decision intelligence — analytics that changes what people do next. Capabilities: Business Intelligence, Advanced Analytics, Customer Intelligence, Marketing Analytics, Audience Segmentation, Decision Intelligence, AI-powered Analytics, Metrics & Semantic Layer.
- [AdTech & MarTech](https://marquedata.com/solutions/adtech-martech/): Customer data platforms, identity resolution and activation that turn data into personalised experiences. Capabilities: Customer Data Platform, Identity Resolution, Customer 360, Audience Segmentation, Customer Intelligence, Campaign Orchestration, Personalization, Marketing Automation, Advertising Activation, Measurement & Attribution.

## Services
- Strategy & Advisory: Data and AI strategy, platform assessments, architecture reviews and use-case prioritisation tied to business value.
- Architecture & Design: Target-state architectures for data platforms, AI systems and customer data — secure, scalable and cost-aware.
- Engineering & Delivery: Hands-on engineering of pipelines, platforms, models and AI applications using modern delivery practices.
- Modernisation & Migration: Move legacy warehouses, ETL and reporting to modern cloud platforms with minimal disruption.
- AI Enablement: Take AI from pilots to production with evaluation, guardrails, LLMOps and change enablement.
- Managed Data & AI Operations: Ongoing operation, monitoring and optimisation of data and AI platforms.

## Products (in development — not yet generally available)
- [AI Campaign Intelligence](https://marquedata.com/products/ai-campaign-intelligence/) — Coming Soon: Understand what is working across campaigns — and what to do next.
- [Customer Intelligence Platform](https://marquedata.com/products/customer-intelligence-platform/) — Coming Soon: Customer 360, segmentation and predictive insight on your own data.
- [Campaign Automation Engine](https://marquedata.com/products/campaign-automation-engine/) — In Research: Signal-driven journeys, orchestrated across channels.
- [Data & AI Platform](https://marquedata.com/products/data-ai-platform/) — In Research: An accelerator for governed lakehouse, analytics and AI.

## Industries
- Financial Services, Insurance, Retail & E-commerce, Healthcare, Telecommunications, Media & Advertising, Technology

## Insights (articles)
- [Choosing a Data Platform in India: Databricks, Snowflake or BigQuery](https://marquedata.com/insights/choosing-a-data-platform-in-india/): How to choose between Databricks, Snowflake, BigQuery and other cloud data platforms in India: workloads, cost, skills, data residency and DPDP Act considerations.
- [LLM Fine-Tuning vs RAG: Which Should You Use?](https://marquedata.com/insights/llm-fine-tuning-vs-rag/): RAG supplies facts at question time; fine-tuning changes how a model behaves. How to choose between them, what fine-tuning needs, and how to prove it worked.
- [Composable CDP: Customer Data Platform Capabilities on Your Lakehouse](https://marquedata.com/insights/composable-cdp-lakehouse/): How to build identity resolution, Customer 360, segmentation and activation on the data platform you already own — and when a packaged CDP is still the better choice.
- [Medallion Architecture on Databricks: A Practical Guide](https://marquedata.com/insights/medallion-architecture-databricks/): How to design bronze, silver and gold layers on Databricks that stay reliable at scale — with patterns for ingestion, data quality, governance and cost.
- [RAG in Production: An Engineering Checklist](https://marquedata.com/insights/rag-in-production-checklist/): What changes when a retrieval-augmented generation prototype becomes a production system — retrieval quality, evaluation, permissions, guardrails, observability and cost.

## Engineering blog (technical posts by our team and the community)
- [LLM Fine-Tuning Explained: From Fundamentals to LoRA, QLoRA, and Production Deployment](https://marquedata.com/blog/llm-fine-tuning-explained-from-fundamentals-to-lora-qlora/): Learn how LLM fine-tuning works, when to use it instead of RAG or prompt engineering, how LoRA and QLoRA reduce training costs, and how to build a production-ready fine-tuning pipeline.
- [CDC into the Lakehouse: Idempotent MERGE, SCD Type 2, Deletes and Out-of-Order Data](https://marquedata.com/blog/cdc-into-the-lakehouse-scd2-delta-merge/): Engineering change data capture on Delta Lake: deduplication, sequence guards, soft deletes, SCD2 in one MERGE, rebuilding history for late data, idempotent streams.
- [Inside the Databricks Lakehouse: Delta Log, Liquid Clustering, Deletion Vectors and Unity Catalog](https://marquedata.com/blog/databricks-lakehouse-internals/): How Delta's transaction log and concurrency work, plus time travel, deletion vectors, liquid clustering, change data feed, Unity Catalog, Auto Loader and Lakeflow.
- [Gradient-Boosted Trees from First Principles: How XGBoost and LightGBM Really Work](https://marquedata.com/blog/gradient-boosted-trees-from-first-principles/): The maths behind gradient boosting, the split-gain formula, histogram and leaf-wise growth, missing values, categoricals, tuning, calibration and SHAP — with code.
- [Kafka Event Processing in Production: Ordering, Exactly-Once, Rebalancing and Failure Handling](https://marquedata.com/blog/kafka-event-processing-in-production/): Partitioning and ordering, durable producers, transactions and exactly-once limits, consumer rebalancing, retries and DLQs, schema evolution and Spark streaming on Kafka.
- [Fine-Tuning LLMs in Depth: LoRA, QLoRA, Training Data, Evaluation and Serving](https://marquedata.com/blog/llm-fine-tuning-lora-qlora-in-depth/): How LoRA and QLoRA work, GPU memory maths, building training data, a working SFT pipeline, hyperparameters, DPO preference tuning, evaluation and multi-adapter serving.
- [A Field Guide to Machine Learning Algorithms: Assumptions, Failure Modes and When We Use Each](https://marquedata.com/blog/machine-learning-algorithms-field-guide/): Linear and logistic models, k-NN, SVMs, random forests, boosting, clustering, anomaly detection, forecasting and recommenders — what each assumes and how each fails.
- [How We Productionise Data and AI on Databricks: Bundles, CI/CD, Testing, MLOps and Operations](https://marquedata.com/blog/productionising-data-and-ai-on-databricks/): Our playbook from notebook to production: repo layout, Asset Bundles, CI/CD with service principals, data testing, idempotent backfills, MLflow and operations.

## Glossary (definitions)
- [Data lakehouse](https://marquedata.com/glossary/data-lakehouse/): A data lakehouse is a data architecture that stores data in open file formats on low-cost cloud object storage, and adds warehouse features on top — ACID transactions, schemas, governance and fast SQL — so data engineering, BI and machine learning work on one copy of the data.
- [Medallion architecture](https://marquedata.com/glossary/medallion-architecture/): Medallion architecture is a way of organising lakehouse data into three layers of increasing quality: bronze holds raw data as ingested, silver holds cleaned and conformed data, and gold holds business-ready tables and aggregates for analytics, reporting and machine learning.
- [Delta Lake](https://marquedata.com/glossary/delta-lake/): Delta Lake is an open-source table format and storage layer that adds ACID transactions, schema enforcement, time travel and efficient upserts (MERGE) to Parquet files on cloud object storage.
- [Unity Catalog](https://marquedata.com/glossary/unity-catalog/): Unity Catalog is the Databricks governance layer for data and AI assets.
- [Apache Spark](https://marquedata.com/glossary/apache-spark/): Apache Spark is an open-source distributed engine for processing large volumes of data in parallel across a cluster.
- [Change data capture (CDC)](https://marquedata.com/glossary/change-data-capture/): Change data capture (CDC) is a technique for detecting inserts, updates and deletes in a source database and delivering only those changes to downstream systems, usually by reading the database transaction log.
- [Slowly changing dimension (SCD Type 2)](https://marquedata.com/glossary/slowly-changing-dimension/): A slowly changing dimension (SCD) is a dimension table whose attributes change occasionally over time.
- [Data mesh](https://marquedata.com/glossary/data-mesh/): Data mesh is an organisational approach to data in which business domains own and publish their data as products, supported by a shared self-service platform and federated governance, instead of routing all data work through one central team.
- [Data governance](https://marquedata.com/glossary/data-governance/): Data governance is the set of policies, roles, processes and tools that decide who can use which data, for what purpose, and to what quality standard.
- [Retrieval-augmented generation (RAG)](https://marquedata.com/glossary/retrieval-augmented-generation/): Retrieval-augmented generation (RAG) is a technique in which a system first retrieves relevant passages from your own documents or data, then gives them to a large language model as context, so answers are grounded in current, private facts and can cite their sources.
- [Vector database](https://marquedata.com/glossary/vector-database/): A vector database stores embeddings — numeric representations of text, images or other data — and finds the items most similar to a query by distance between vectors.
- [Agentic AI](https://marquedata.com/glossary/agentic-ai/): Agentic AI refers to AI systems that pursue a goal over several steps: a language model plans, calls tools or APIs, observes the results and decides what to do next, within guardrails set by people.
- [LLM fine-tuning](https://marquedata.com/glossary/llm-fine-tuning/): LLM fine-tuning is continued training of a pre-trained large language model on your own examples, so it learns a specific behaviour — an output format, a classification, a tone or a specialised task.
- [LoRA and QLoRA](https://marquedata.com/glossary/lora/): LoRA (low-rank adaptation) fine-tunes a large model by freezing its original weights and training small low-rank matrices added to selected layers.
- [MLOps](https://marquedata.com/glossary/mlops/): MLOps (machine learning operations) is the practice of building, deploying, monitoring and retraining machine learning models reliably in production.
- [Decision intelligence](https://marquedata.com/glossary/decision-intelligence/): Decision intelligence is a discipline that combines data, analytics, machine learning and business rules to improve specific decisions — not just report on the past.
- [Customer data platform (CDP)](https://marquedata.com/glossary/customer-data-platform/): A customer data platform (CDP) collects customer data from every channel, resolves it into unified profiles, builds audiences and segments, and sends them to marketing, advertising and service tools.
- [Composable CDP](https://marquedata.com/glossary/composable-cdp/): A composable CDP delivers customer data platform capabilities — identity resolution, Customer 360, segmentation and activation — directly on an organisation's existing data warehouse or lakehouse, instead of copying customer data into a separate packaged CDP.
- [Identity resolution](https://marquedata.com/glossary/identity-resolution/): Identity resolution is the process of matching records that belong to the same person or household across devices, channels and systems — using deterministic keys such as email or customer ID, and probabilistic signals — to build one unified customer profile.
- [Reverse ETL](https://marquedata.com/glossary/reverse-etl/): Reverse ETL copies modelled data from a data warehouse or lakehouse back into operational tools — CRM, marketing automation, ad platforms, support desks — so teams can act on segments, scores and metrics inside the systems they already use.

## Key pages
- [Home](https://marquedata.com/)
- [About](https://marquedata.com/about/)
- [Technology](https://marquedata.com/technology/)
- [Industries](https://marquedata.com/industries/)
- [Services](https://marquedata.com/services/)
- [Insights](https://marquedata.com/insights/)
- [Engineering blog](https://marquedata.com/blog/)
- [Glossary](https://marquedata.com/glossary/)
- [Contact](https://marquedata.com/contact/)
- [Full text of all articles and definitions](https://marquedata.com/llms-full.txt)
- [RSS feed](https://marquedata.com/rss.xml)

## FAQ
### What is Marquedata?
Marquedata (marquedata.com) is an independent data and AI engineering company based in Kolkata, India. It builds cloud data platforms, data science and machine learning (model training and LLM fine-tuning), analytics, generative and agentic AI, and customer intelligence for AdTech and MarTech, and is developing its own B2B SaaS products. It is not affiliated with data-storage hardware or other similarly named businesses.

### Does Marquedata provide data engineering and AI services in India?
Yes. Marquedata is based in Kolkata and works with organisations in India and worldwide on data engineering, cloud data platforms (Databricks, Snowflake, AWS, Azure, Google Cloud), data science and machine learning (ML model training, LLM fine-tuning, MLOps), generative AI and agentic AI, analytics and decision intelligence, and AdTech and MarTech data solutions. Start with a discovery conversation at info@marquedata.com.

### What does Marquedata do?
Marquedata is a data and AI engineering company. We design and build modern data platforms, data science and machine learning solutions (including model training and LLM fine-tuning), analytics and decision intelligence, generative and agentic AI solutions, and customer intelligence systems for AdTech and MarTech — and we are developing our own SaaS products from the patterns we deliver.

### Which technologies does Marquedata work with?
Our core stack includes Databricks, Apache Spark, Delta Lake, Kafka, Snowflake and SQL on AWS, Azure and Google Cloud, plus large language models, retrieval-augmented generation (RAG), vector search and AI agent frameworks. For marketing we integrate with Meta Ads, Google Ads, Adobe and marketing automation platforms.

### How does a typical engagement start?
Engagements start with a short discovery to understand goals, the current data landscape and constraints, and to agree measurable outcomes. We then design the architecture and deliver in small increments of working software, before scaling, handing over or operating the solution.

### Are Marquedata products available today?
Not yet. AI Campaign Intelligence, Customer Intelligence Platform, Campaign Automation Engine and a Data & AI Platform accelerator are in research or development. Organisations can join early access to become design partners and shape the roadmap.

### Where is Marquedata based?
Marquedata is based in Kolkata, India, and works with clients globally. You can reach the team at info@marquedata.com or through the contact page.

### What is LLM fine-tuning and when is it worth doing?
LLM fine-tuning continues training a pre-trained language model on your own examples so it learns your terminology, format or task. It is worth it when prompting and retrieval are not enough, such as for consistent style, specialised classification or lower-latency smaller models. We always measure gains against a held-out evaluation set first.

### Should we fine-tune a model or use RAG?
They solve different problems. RAG supplies up-to-date facts from your documents at question time; fine-tuning changes how the model behaves, writes or classifies. Many production systems combine both. We start with the simplest approach that meets the quality target and add fine-tuning only where evaluation shows a clear benefit.

### What do you need to train a machine learning model?
A clearly defined prediction target, enough representative historical data, a way to measure success in business terms and a path to deploy the result. We help assess data readiness, build features and training pipelines, evaluate models against baselines, and deploy them with monitoring.

### What is MLOps?
MLOps applies engineering discipline to machine learning: versioned data and code, experiment tracking, a model registry, automated testing and deployment, and monitoring for drift and quality, with retraining when performance drops. It keeps models reproducible, auditable and reliable after the first release.

### What is a lakehouse and why use one?
A lakehouse combines the low-cost, flexible storage of a data lake with the reliability and performance of a data warehouse, using open table formats such as Delta Lake. It lets BI, machine learning and generative AI work from one governed copy of data instead of several disconnected systems.

### What is the medallion architecture?
The medallion architecture organises data in three layers: bronze holds raw, replayable source data; silver holds cleansed, validated and conformed data; gold holds business-ready, modelled data for analytics and AI. Each layer adds quality checks, making issues easier to trace and fix.

### Do you work on Databricks, Snowflake or both?
Both. We recommend a platform based on your workloads, existing skills, cloud and cost profile. Databricks suits heavy data engineering, streaming and machine learning; Snowflake suits SQL-centric warehousing. Many organisations use them together.

### How do you handle data quality and governance?
Quality and governance are built into pipelines from the start: automated expectations and quarantine at each layer, access control and classification through a catalogue such as Unity Catalog, and end-to-end lineage and observability so teams can see freshness, volume and schema changes.

### What is retrieval-augmented generation (RAG)?
RAG is a pattern where a large language model answers using documents and data retrieved from your own knowledge sources at question time. It grounds answers in trusted enterprise content, supports citations and respects permissions, reducing hallucinations compared with a model answering from its training data alone.

### What is agentic AI?
Agentic AI uses one or more AI agents that plan and take multi-step actions — calling tools, APIs and data sources — to complete a task. In enterprise settings we design agents with guardrails, audit trails and human-in-the-loop approval for decisions that matter.

### How do you move a GenAI proof of concept into production?
Production readiness adds systematic evaluation of answer quality, permission-aware retrieval, guardrails, observability, cost controls and LLMOps for versioning and monitoring. We treat these as engineering requirements from the start rather than steps added after a demo.

### Which language models do you use?
We are model-agnostic and select models per use case based on quality, latency, cost, data residency and security requirements, using commercial APIs or open models hosted in your cloud.

### What is decision intelligence?
Decision intelligence connects data, analytics and AI models to specific business decisions, so insight becomes a recommendation or an automated action rather than only a dashboard. It focuses on which decisions matter, the data they need and how outcomes are measured.

### What is a semantic layer?
A semantic layer defines business metrics and entities once — such as revenue, active customer or churn — and makes those definitions reusable across BI tools, notebooks and AI assistants, so every team reports the same numbers.

### Can business users ask questions of data in plain language?
Yes. AI-powered analytics lets users ask questions in natural language and receive answers generated from governed metrics and data, with the underlying query visible for trust and review.

### What is a composable CDP?
A composable customer data platform builds CDP capabilities — identity resolution, unified profiles, segmentation and activation — directly on your existing data warehouse or lakehouse, instead of copying customer data into a separate packaged CDP. It reduces data silos and keeps governance in one place.

### What is identity resolution?
Identity resolution links records from web, mobile, CRM, transactions and marketing systems that belong to the same person or household, using deterministic matches such as email or login and, where appropriate, probabilistic signals. It is the foundation of a Customer 360 view.

### Which advertising and marketing platforms can audiences be activated to?
Audiences can be activated to advertising platforms such as Meta Ads and Google Ads, the Adobe Experience Cloud ecosystem, and CRM, email and marketing automation tools, with results fed back for closed-loop measurement.