What's your day-to-day tech stack?

LangGraph and LangChain for orchestration, Azure OpenAI and Azure AI Foundry for models, Pinecone and Azure AI Search for retrieval, Snowflake and Databricks on the data side, FastAPI for serving, and MCP for tool integration. I work across the full LLM lifecycle — prompting, RAG, fine-tuning with LoRA, agent orchestration, evaluation, and LLMOps.

Are you available right now?

Yes — I'm open to full-time roles, contract projects, and freelance consulting. The fastest way to reach me is booking a quick call at cal.com/swarnava-dutta/quick-call, or connect on LinkedIn, or email swarnava.dev@gmail.com. I usually reply within 12 hours.

What does working with you look like?

I keep it to five steps: first we get clear on the problem and what success measurably means, then I design the architecture — retrieval, agents, models, guardrails. I ship a working end-to-end slice early and demo every week, build evaluation and observability before launch, and hand off with runbooks so your team owns it without me. No long silent build phases, no big-bang launch surprises.

Do you work with remote or international teams?

Yes — most of my work has been with distributed teams across the US and Europe, so collaborating across time zones is the norm for me. I'm based in Kolkata, India (IST). Contracts, full-time remote roles, and hybrid arrangements within India all work.

Available for New Opportunities

Kolkata, India

Hi, I'mSwarnava DuttaLead AI Architect

Q: Who is Swarnava Dutta?

Swarnava Dutta is a Lead AI Architect based in Kolkata, India, with 6 years of experience in generative AI engineering. He designs and ships production-grade AI systems — multi-agent workflows built with LangGraph, retrieval-augmented generation (RAG) architectures, and enterprise LLM platforms on Azure OpenAI and Azure AI Foundry. His most recent project is LindsAI, a multi-agent system with tool calling into JIRA and SharePoint plus dynamic web search. He currently works at TEKsystems as a Senior GenAI Engineer and Technical Lead.

Q: What kind of work can I hire you for?

Four things, mainly: designing multi-agent AI systems (LangGraph, MCP, tool calling into platforms like JIRA and SharePoint, guardrails), building RAG architecture and vector search that actually retrieves the right things (Azure AI Search, Pinecone, hybrid search, reranking, RAGAS evaluation), LLM fine-tuning and optimization (LoRA/PEFT, dataset design, inference tuning), and enterprise AI architecture that takes teams from pilot to a secure, observable production platform. If your problem doesn't fit one of those boxes neatly, ask anyway — most interesting problems don't.

I design agentic workflows, production RAG pipelines, and LLMOps infrastructure enterprise AI systems built to hold up under real traffic, real users, and real compliance pressure.

Multi-Agent SystemsFine-tuning & LoRARAG ArchitectureMCP Workflows

Open to

Full-time OpportunitiesContract ProjectsFreelance Consulting

Download Resume

system blueprint

agentic workflow

Incoming request

Planner agent

LangGraph

Breaks the request into steps and routes work to specialists

RAG retrieval

Azure AI Search · Pinecone

MCP tool calls

JIRA · SharePoint · Web

Evaluation gate

Human-in-the-loop

RAGAS scoring and an approval checkpoint before anything moves forward

Ship to production

Monitored, logged, and easy for teams to run

live

The agentic pattern behind my recent builds — planning, grounded retrieval, tool execution, and a human gate before production.

Trusted by teams

Work that had to hold up for real teams and real workflows.

High-compliance AI systems for regulated sectors — global banking, healthcare, enterprise platforms, and AI-native products — where "it mostly works" isn't good enough.

Where it shows up

BankingHealthcareEnterprise platformsRetailRecruitment

The common thread is building systems people can actually use day to day, not just demo once and forget.

MUFG Bank

Solventum

Google

WhisperIt

3M Health

Brand platforms delivered through employer and consulting engagements at TEKsystems, Syren Technologies, RWS Moravia, and Capgemini, alongside direct product work.

Delivery record

Production outcomes, not demo numbers.

Every figure below is sourced from a specific role or system in the experience timeline further down this page.

Systems delivered

Enterprise AI systems shipped end to end

Daily queries

10K+

Scaled in production on a supply-chain copilot

Production uptime

99.9%

Maintained across live enterprise workloads

Efficiency gain

67%

Reduction in manual operational workloads

Certifications

Verified cloud, data, and AI credentials

Enterprise reach

10+

Organizations and teams supported

Professional experience

The path behind how I build production AI today.

Across 5+ years

The work moved from data platforms into search, copilots, and enterprise GenAI systems, with reliability and delivery pressure increasing at each step.

Data engineeringEnterprise searchGenAI deliveryTechnical leadership

Dec 2025 - Present

TEKsystems Global Services Pvt. Ltd

Senior Generative AI Engineer

Technical Lead

Dec 2025 - Present

Cut intake-to-ticket turnaround from 2+ days to under 10 minutes, compressing P95 pipeline latency by 85% with parallelized Redis-backed workers

Improved learner skill-gap detection accuracy by 37% across 6K+ monthly sessions with adaptive quiz generation and continuous RAGAS evaluation

LangGraphMCPAzure AI FoundryAzure AI SearchRedisRAGASFastAPI

Jul 2025 - Dec 2025

Syren Technologies Private Limited

Senior Generative AI Engineer

Technical Lead

Jul 2025 - Dec 2025

Scaled a supply chain AI copilot to 10K+ daily queries, reducing latency by 63% while maintaining 99.9% uptime

Boosted contextual recall by 34%, improved multi-turn response accuracy by 28%, and reduced context errors by 42%

LangGraphPineconeFastAPIMistral-7BLoRAMCPBLEU/ROUGE-L

Jan 2024 - Jul 2025

RWS Moravia India Private Limited

Senior Generative AI Engineer

Technical Lead

Jan 2024 - Jul 2025

Launched the company's first production-grade multilingual enterprise RAG platform and AI agent

Cut manual knowledge lookup by 43% and established rigorous BLEU and NDCG evaluation benchmarks

RAGMultilingual AILangChainAzure OpenAIBLEUNDCG

Jul 2022 - Dec 2023

BinaryERP Software Private Limited

Junior Data Scientist

Jul 2022 - Dec 2023

Built semantic search prototypes that drove adoption across 13 B2B users

Improved dataset usability by 26% and laid the foundation for enterprise NLP initiatives

PythonNLPEmbeddingsText ClassificationSemantic SearchData Pipelines

Jan 2021 - Jul 2022

Capgemini Technology Services India Limited

Associate Data Engineer

Jan 2021 - Jul 2022

Reduced nightly batch runtime by 95% for a global banking client

Decreased batch failures by 61% while preserving SLA compliance and improving ETL validation automation

Azure Data LakeAzure Data FactorySQL ServerETLData ValidationAutomation

Download Resume

Enterprise AI deployments

Case studies where the system had to help a real team.

What they have in common

Every system here started from a clear user task, shipped with metrics tied to adoption or efficiency, and held up under daily production use.

Clear user taskMeasurable impactSurvives daily use

Workflow Automation01

LindsAI

Multi-agent automation with JIRA, SharePoint & live web search

98%

intent accuracy

<10 min

intake to ticket

View case studyLangGraph +3

Revenue Enablement02

GRIT

Adaptive sales enablement for regulated medical teams

65%

faster onboarding

50K+

monthly sessions

View case studyRAG +3

Voice Workflow03

Talksmith

Voice-first analytics for teams that live in data

200+

daily users

business teams

View case studyLangGraph +3

Interview Prep04

IntervueRecall

Interview recordings turned into clean question lists

1-click

audio upload

extracted list

View case studyAudio Transcription +3

Knowledge Copilot05

InsightDesk AI

A knowledge copilot for documents, tables, and daily decisions

10K+

monthly questions

~50%

analyst lift

View case studyDatabricks +3

Content Operations06

LinkedInfluencer

AI content intelligence for building a credible LinkedIn voice

20+ hrs

saved weekly

A/B

content optimization

View case studyLangGraph +3

Talent Screening07

FitScout

Recruiter intelligence for faster resume screening

1000+

resumes monthly

90%+

screening accuracy

View case studyLlamaIndex +3

Legal Ops08

WhisperIt Legal Copilot

A private document assistant for faster legal review workflows

45%

faster turnaround

GDPR

privacy aligned

View case studyRAG +3

Shared Services09

OpsAnswer Hub

One assistant for HR, finance, policy, and payroll questions

40%

routine query share

200+ hrs

saved monthly

View case studyLangChain +3

Where I help

The work usually starts in one of these lanes.

What usually matters

Some teams need a stronger retrieval core, others need multi-agent workflows or a safer path from pilot to platform. These are the patterns I most often design around.

Clear scopeEvaluation earlyReliable handoff

Complex workflow orchestration01

Multi-Agent Systems

Design multi-agent systems that coordinate tools, memory, approvals, and specialist roles without becoming hard to manage.

Best suited for

Teams replacing repetitive manual work with planner, reviewer, and executor workflows.

Clear agent roles and handoffs

Human review steps and safety rails

LangGraphMCPTool callingGuardrails

Ground models in your knowledge02

RAG Architecture & Vector Search

Build retrieval systems that surface the right context quickly, keep answers grounded, and make knowledge easier to use.

Best suited for

Products and internal copilots that need trusted answers from documents, data systems, and live business context.

Ingestion, chunking, and indexing plan

Hybrid retrieval with reranking and citations

Azure AI SearchPineconeRerankingRAG evals

Sharpen models for domain fit03

Model Fine-tuning & Optimization

Improve quality, speed, and consistency with the right mix of fine-tuning, prompt shaping, dataset design, and inference optimization.

Best suited for

Teams that already see promise but need better domain accuracy, tone control, or task fit.

Dataset curation and benchmark framing

LoRA or PEFT fine-tuning workflow design

LoRAPEFTBenchmarksInference tuning

Turn pilots into platforms04

Enterprise AI Architecture

Shape the end-to-end AI setup so it is secure, observable, and easy for real teams to run over time.

Best suited for

Teams moving from early AI experiments to dependable platforms, governance, and long-term use.

A reference setup for services and data flow

Security, compliance, and deployment patterns

AzureKubernetesCI/CDMonitoring

Engagement models

Concrete scopes, timelines, and deliverables.

Every engagement starts from one of these formats. Baseline rates and a delivery estimate are shared in the first scoping call.

1-2 weeks · Fixed scope

AI Architecture Audit

Strong foundations in Python, ML tooling, and data platforms that support dependable delivery.

Python and MLData engineeringAnalytics workflows

Often includes

PythonPyTorchTensorFlowSQLPandasNumPyPySparkTypeScript/ReactKafkaData ModelingUnity CatalogDelta Lake

Credentials

The learning layer behind the delivery work.

Built around practice

DAG design, scheduling, and workflow orchestration fundamentals.

Verify credential

Where it started

Bachelor of Technology, Computer Science and Engineering

Maulana Abul Kalam Azad University of Technology

2016–2020

How I work

A process built around shipping, not slideware.

Why I work this way

I've watched too many AI projects die in a six-month build phase nobody saw. So every step here produces something you can see and test — and the system gets stronger every week.

Weekly demosEvaluation earlyDocumented handoff

1
Discover & scope
I start by understanding your workflow, your users, and your data — and what success would actually look like. No code until that's clear.
Problem definitionSuccess metricsScope document
2
Design the architecture
Then I pick the retrieval strategy, agent topology, models, and guardrails that fit your constraints — not whatever is trending that week.
Architecture diagramModel selectionRisk review
3
Build in thin slices
I ship a working end-to-end slice early, then iterate against real data and real user feedback. You see progress every week, not at the end.
Working prototypeWeekly demosIteration log
4
Evaluate & harden
Before launch I build eval pipelines, tune latency and cost, and make sure failures show up on a dashboard — not in a user complaint.
Eval suiteLatency budgetObservability
5
Launch & hand off
I deploy with monitoring in place, write the runbooks, and make sure your team can run the system confidently without me.
Production deployRunbooksTeam enablement

Professional profile

I build AI systems that teams can still trust after launch.

I'm an enterprise AI architect based in Kolkata, India. For 5+ years I've worked across data engineering, search, and generative AI, building systems where answer quality, reliability, and business value all matter at the same time.

Multi-Agent SystemsFine-tuning & LoRARAG ArchitectureMCP WorkflowsEvaluation Pipelines

What I focus on

As a Lead AI Architect, I care about AI systems that keep working once teams start relying on them. That mindset comes from my years in cloud data platforms, automation, and reliability-focused engineering.

Much of my recent work centers on LangGraph orchestration, RAG, and domain-specific AI products. Recent platforms have supported 10K+ daily queries while maintaining 99.9% uptime in live environments.

I engineer for outcomes teams can measure: lower latency, automated workflows, and faster decisions. That focus has delivered efficiency gains as high as 67% while keeping each system clear enough for teams to maintain after launch.

How I work

The things I care about when building AI systems.

Quality

Measure the behavior before you trust the behavior.

Strong AI systems earn trust through clear results and careful testing. I set up evaluation early so teams can see what is working and improve with confidence.

Retrieval quality is checked before scale becomes the priority.

Answer quality is reviewed with clear task-based checks.

Decisions stay tied to user needs and business goals.

Reliability

Build for everyday use, not only for demos.

The best AI products keep working when data is messy, traffic grows, and more than one team needs to use or support them.

Workflows are kept clear before complexity piles up.

Latency, fallbacks, and caching are planned from the start.

Monitoring and team handoff are considered alongside model behavior.

Usefulness

Make the result useful enough that people keep coming back.

The most valuable AI work saves time, improves decisions, or removes repetitive effort. That is the standard I like to build toward.

The business goal stays visible from discovery through launch.

Security, guardrails, and handoff readiness are built in early.

The final system should feel easy to use, extend, and trust.

Quick answers

Things people usually ask me.

Something else on your mind?

If your question isn't here, just ask — a quick call or a short email works. I read everything myself and usually reply within a day.

Direct answersNo sales pitch24h response

Swarnava Dutta is a Lead AI Architect based in Kolkata, India, with 6 years of experience in generative AI engineering. He designs and ships production-grade AI systems — multi-agent workflows built with LangGraph, retrieval-augmented generation (RAG) architectures, and enterprise LLM platforms on Azure OpenAI and Azure AI Foundry. His most recent project is LindsAI, a multi-agent system with tool calling into JIRA and SharePoint plus dynamic web search. He currently works at TEKsystems as a Senior GenAI Engineer and Technical Lead.

Four things, mainly: designing multi-agent AI systems (LangGraph, MCP, tool calling into platforms like JIRA and SharePoint, guardrails), building RAG architecture and vector search that actually retrieves the right things (Azure AI Search, Pinecone, hybrid search, reranking, RAGAS evaluation), LLM fine-tuning and optimization (LoRA/PEFT, dataset design, inference tuning), and enterprise AI architecture that takes teams from pilot to a secure, observable production platform. If your problem doesn't fit one of those boxes neatly, ask anyway — most interesting problems don't.

Next step

If the fit feels right, we can keep the first step simple.

Start with the resume, look through the work, or send a quick note first. Whatever helps you get context fastest.

RolesConsultingArchitecture reviewsWithin 12 hours

Download Resume

Let's talk

If there's a role, build, or idea worth exploring, send a note.

Best starting points

Bring a consulting brief, a full-time opportunity, an architecture question, or a specific AI system that needs clearer shape.

Full-time OpportunitiesContract ProjectsFreelance Consulting

Calendar not loading? Open the booking page in a new tab.

Project brief

Send a note

Share the problem, timeline, or role you have in mind. A short brief is enough.

swarnava.dev@gmail.com

Location

Kolkata, India (IST)

Response time

Within 12 hours

Connect

From the blog

Notes from building production AI.

Illustration of how to install flash attention 3: A wide workbench: left side, a GPU chip on a jig checked with calipers

Nvidia H100

September 6, 20269 min read

How to Install FlashAttention-3 on NVIDIA Hopper GPUs

Learn how to install Flash Attention 3 on NVIDIA Hopper GPUs, match CUDA and PyTorch, build from source, verify the package, and fix common errors.

Illustration of how to use qlora huggingface: A massive stone block on the left is compressed by a wide press into a small

Hugging Face Peft

September 5, 20268 min read

How to Use QLoRA Hugging Face on a Single GPU

Learn how to use QLoRA Hugging Face to fine-tune Llama 3 in 4-bit NF4 with PEFT and TRL, then compare VRAM use against LoRA and DoRA adapters.

Illustration of flash attention vs grouped query attention: Wide workbench: left side, a narrow water pipe with a tight

Flashattention-2

September 4, 20268 min read

Flash Attention vs Grouped Query Attention in PyTorch

Learn how Flash Attention vs grouped query attention affects latency, VRAM, and KV cache size, with PyTorch benchmarks and implementation guidance.

Illustration of how to implement paged attention: A long wall of uniform lockers spans the frame; on the left, mismatched

Vllm

September 3, 20268 min read

How to Implement Paged Attention in PyTorch and vLLM

Learn how to implement paged attention in PyTorch with a block KV cache, decoding tests, memory math, benchmarks, and a clear mapping to vLLM.

View all posts

Hi, I'mSwarnava DuttaLead AI Architect

Work that had to hold up for real teams and real workflows.

Production outcomes, not demo numbers.

The path behind how I build production AI today.

Senior Generative AI Engineer

Senior Generative AI Engineer

Senior Generative AI Engineer

Junior Data Scientist

Associate Data Engineer

Case studies where the system had to help a real team.

LindsAI

GRIT

Talksmith

IntervueRecall

InsightDesk AI

LinkedInfluencer

FitScout

WhisperIt Legal Copilot

OpsAnswer Hub

The work usually starts in one of these lanes.

Multi-Agent Systems

RAG Architecture & Vector Search

Model Fine-tuning & Optimization

Enterprise AI Architecture

Concrete scopes, timelines, and deliverables.

AI Architecture Audit

Discovery Sprint

End-to-End Delivery

Fractional AI Architect

Production-grade AI tooling and LLMOps frameworks.

GenAI & LLMs

Azure AI

Evaluation & Quality

Vector Databases

LLMOps

Programming & Data

The learning layer behind the delivery work.

Azure Fundamentals

Azure Data Fundamentals

Databricks Fundamentals

Generative AI Fundamentals

Azure Databricks Platform Architect

Apache Airflow Fundamentals

Bachelor of Technology, Computer Science and Engineering

A process built around shipping, not slideware.

Discover & scope

Design the architecture

Build in thin slices

Evaluate & harden

Launch & hand off

I build AI systems that teams can still trust after launch.

The things I care about when building AI systems.

Measure the behavior before you trust the behavior.

Build for everyday use, not only for demos.

Make the result useful enough that people keep coming back.

Things people usually ask me.

Who is Swarnava Dutta?

What kind of work can I hire you for?

What's your day-to-day tech stack?

Are you available right now?

What does working with you look like?

Do you work with remote or international teams?

If the fit feels right, we can keep the first step simple.

If there's a role, build, or idea worth exploring, send a note.

Send a note

Notes from building production AI.

How to Install FlashAttention-3 on NVIDIA Hopper GPUs

How to Use QLoRA Hugging Face on a Single GPU

Flash Attention vs Grouped Query Attention in PyTorch

How to Implement Paged Attention in PyTorch and vLLM

Work that had to hold up for real teams and real workflows.

Production outcomes, not demo numbers.

The path behind how I build production AI today.

Senior Generative AI Engineer

Senior Generative AI Engineer

Senior Generative AI Engineer

Junior Data Scientist

Associate Data Engineer

Case studies where the system had to help a real team.

LindsAI