Open to Summer / Full-time roles in AI Product & Applied Science

Building autonomous AI
that ships to production.

I am a ML Scientist

9+ years turning frontier research into 39 end-to-end AI systems across Generative AI, Agentic AI, LLMs, NLP and Vision — at KLA, Amazon, Oracle, Rakuten & Cisco. Now fusing deep technical craft with a STEM MBA to scale AI into business value.

0
Years in AI/ML
0
End-to-end AI systems shipped
0
Customers served by my models
0
Autonomous PRs by AURORA
0
MiniTV users experience improved
0
Engineers mentored across 7 teams
0
ARR portfolio owned at Oracle
0
Enterprise clients acquired
01 — Profile

Where deep research meets business impact.

I am an ML Scientist and technical leader with 9+ years spent at the frontier — shipping 39 highly scalable, end-to-end AI systems from research to production. My work spans Generative AI, Agentic AI, multimodal LLMs, NLP/NLU, computer vision, recommendation and reinforcement learning.

At KLA I built AURORA, a 10-agent autonomous system that turns bug reports and feature requests into production-ready pull requests in ~6 minutes — shipping 120+ PRs across 9 repositories with zero human intervention. At Amazon, as Principal Applied Scientist, I led a 15-person cross-functional team to build LLM-powered personalization that lifted CTR 44% and dwell time 53%, and authored the science roadmap that VPs approved for 2024.

I don't just build models — I build teams and translate ambiguity into strategy. I've mentored 45+ engineers, secured VP-level buy-in for 12 strategic initiatives, and owned multi-quarter roadmaps across Retail, Devices, Prime Video and AWS. My STEM MBA at Ohio State (Fisher Leadership Fellow) sharpens the bridge between technical depth and P&L outcomes.

Underneath it all is a lifelong obsession with excellence: Rank 1 out of 1 million in my 12th board exams, Rank 1 in Computer Science at BITS Pilani, and top-1% national finishes in Physics, Chemistry, Math and Astronomy olympiads.

02 — Experience

A decade shipping AI at global scale.

From humanoid brain-computer interfaces to autonomous multi-agent engineering systems — across six world-class organizations.

AI Scientist Intern · KLA
Ann Arbor, Michigan, USA
May 2026 – Aug 2026
  • Built AURORA — an autonomous 10-agent system converting bug reports & feature requests into production-ready PRs in ~6 min with self-correcting review loops; 120+ PRs across 9 repos, zero human intervention.
  • Developed GAIA & EVA, agentic RAG chatbots (BM25 + MPNet + semantic re-ranking, Gemma dual-model routing) indexing 218 pages into 3,180 chunks at <50ms P99, with streaming & voice — deployed at $0 infra cost.
  • Created WAALI — an NL-driven autonomous wiki-edit agent with explainable-AI confidence scoring, NER, semantic coherence & multi-page batch editing across 12 releases.
  • Designed production label-based RBAC protecting 250K+ nodes across 5 Neo4j DBs — 500+ custom roles, 250 automated tests (100% pass), O(1) bitmap checks, 92% lower credential-leak exposure.
  • Found & remediated 12 critical credential exposures during an ADO→GitHub wiki migration (75 pages, 84 attachments); built an automated secret-redaction & syntax-conversion pipeline.
  • Published 4 Claude Code skills to KLA's internal AI registry (citizen-dev-review, citizen-dev-setup, sage-onboarding, wiki-setup).
Multi-AgentLangGraphRAGNeo4j RBACExplainable AIGemma
Principal Applied Scientist / Senior Tech Lead · Amazon
Bangalore, India
Feb 2021 – Aug 2025
  • Defined the 2024 annual strategic roadmap across 4 orgs (Retail, Devices, Prime Video, AWS) — authored 45 business cases, secured VP approval for 12 initiatives.
  • Led a cross-functional team of 15 to ship Rec-Rainbow, an LLM-powered zero-shot personalization agent — +44% CTR, +53% dwell time, +16% revenue.
  • Accelerated multi-hop AWS Bedrock workflows (Anthropic MCP, LangGraph, CrewAI); built an agent ecosystem (Zephyr, Mixtral-8x7B, WizardCoder, NexusRaven-V2) automating queries for 70M Amazon Pay customers and cutting OpEx 3×.
  • Built a RAG system for AWS Documentation (Pinecone + Longformer + Flan-T5) reaching 87% BertScore, 0.91 MRR.
  • Filed a patent for a CV model — 98% precision / 99% recall, beating VGG/ResNet/Inception/EfficientNet with 9× speed & 14.5× less memory.
  • Evolved Sherlock (BERT + ELECTRA, few-shot) across 6 content types, cutting onboarding from 6 months to 3 weeks; shipped a BART tag-generation model at 93% precision@10, 0.85 MRR.
  • Optimized model execution 250% on EMR, enabling 0.3M batch inferences/day; hierarchical clustering (ResNet + Graph Transformer) hit 99.6% recall / 98.7% precision; MiniTV RecSys reached 0.87 NDCG / 0.89 mAP.
  • Integrated a Mixtral-8x7B multi-agent pipeline on AWS Lex — +31% query-resolution speed, automating 38K support tickets/quarter; lifted Kindle purchases 15% and MiniTV engagement 21% for 250M users.
  • Mentored 45+ engineers across 7 teams; ran biweekly ML sessions and set the org's science vision.
LLMRecSysBedrockMCPPatentNLPLeadership
Senior Applied Scientist / Product Lead · Oracle
Bengaluru, India
Jan 2019 – Jan 2021
  • Pioneered the vision for a Multimodal LLM and a 4-quarter roadmap delivering 25% faster time-to-delivery (release lead time 16→7 weeks).
  • Owned 9 document-understanding models, leading a 16-member team serving a $20M ARR portfolio and acquiring 42 international clients.
  • Designed org-level Knowledge Graphs to automate KTLO on-call, freeing 50+ support engineers.
  • Drove ML lifecycle across 4 Cerner Health verticals impacting 27K customers and lifting retention 18%.
Multimodal LLMKnowledge GraphsDoc AIProduct Strategy
Research Scientist · Rakuten R&D
Bangalore, India
Aug 2018 – Jan 2019
  • Built a YOLOv3 object-detection model to localize antenna-pole structures from video feeds — 2.4× revenue growth KPI, 90% manual-inspection cost eliminated (conf. 0.84, IoU 0.72, deployed on GCP).
Computer VisionYOLOv3GCP
Software Development Engineer · Cisco Systems
Bangalore, India
Jan 2017 – Jul 2018
  • Built a predictive-analytics framework forecasting network failures in real time for Cisco Webex — +78% uptime, $1.8M/yr saved, preventing global outages.
  • Created a lightweight online anomaly-detection framework for device health at 98.4% accuracy; DRL fault-tolerance for Smart Router (87.86% conf.).
Predictive AnalyticsAnomaly DetectionDeep RL
Research Fellow · Microsoft
Bengaluru, India
Jan 2016 – Dec 2016
  • Built a real-time Brain-Computer Interface classifying target stimulus from EEG signals, integrated with a NAO humanoid via Choregraphe — applied to ADHD & motor-activity therapy.
BCI / EEGRoboticsHealthcare AI
03 — Featured Work

Flagship systems, real outcomes.

Ten production systems I architected and shipped end-to-end — from the research and the model to the infra, the tests and the P&L impact. Every metric below is drawn straight from delivery.

🤖
AURORA
Autonomous 10-Agent Engineering System · KLA

A self-correcting, multi-agent software factory that ingests bug reports & feature requests and returns production-ready pull requests — orchestrating triage, routing, code generation, review and verification loops with zero human in the loop.

  • Ten specialized agents (planner, router, coder, reviewer, verifier & more) coordinate over a shared state graph with automatic retries and rollback.
  • Repo-aware verification runs the real build & test suite before a PR is ever opened, keeping merge quality high.
  • Turns a multi-hour engineering task into a ~6-minute autonomous cycle.
~6 minreport → PR
120+PRs shipped
9repos, 0 humans
Multi-AgentSelf-CorrectionCI/CDAutonomy
STACKLangGraph · CrewAI · Claude & Gemma · GitHub API · Python · async orchestration
💬
GAIA & EVA
Agentic RAG Chatbots · KLA

Twin agentic-RAG assistants with hybrid retrieval (BM25 + MPNet dense embeddings + semantic re-ranking) and a Gemma dual-model router, delivering streaming answers with voice input — engineered for near-zero latency at zero infrastructure cost.

  • Indexed 218 pages into 3,180 semantically-chunked passages with metadata-aware retrieval.
  • Dual-model routing sends simple turns to a small model and hard turns to a larger one — protecting both latency and cost.
  • Streaming token output + voice input for a natural, real-time assistant experience.
<50msP99 latency
3,180indexed chunks
$0infra cost
RAGHybrid RetrievalVoiceStreaming
STACKBM25 · MPNet · cross-encoder re-ranker · Gemma · VectorDB · FastAPI
✍️
WAALI
Wiki Agentic AI Live Interface · KLA

A natural-language wiki-editing agent: describe the change in plain English and it plans, drafts and applies it — with explainable-AI confidence scoring, NER extraction, semantic-coherence checks, multi-page batch editing and inline diff previews before anything is committed.

  • Explainable-AI confidence score on every proposed edit so reviewers know what to trust.
  • Batch-edits across many pages in a single instruction while preserving semantic coherence.
  • Inline diff preview keeps a human approval gate on autonomous changes.
12feature releases
XAIconfidence scoring
Batchmulti-page edits
Explainable AINERAgenticDiff Preview
STACKLLM agents · spaCy NER · semantic similarity · Python · Git integration
🔐
Neo4j Label-Based RBAC
Graph Security at Scale · KLA

Production-grade, label-based role-based access control securing a knowledge graph across five databases — 500+ custom roles enforced with O(1) bitmap authorization that holds at unbounded scale, backed by a full automated test suite.

  • Constant-time O(1) bitmap permission checks — authorization cost stays flat as the graph grows.
  • Impersonation & full audit trail for compliance-grade traceability.
  • 250 automated tests at a 100% pass rate guard every role and edge case.
250K+nodes protected
500+custom roles
-92%leak exposure
Neo4jSecurityO(1) BitmapAudit
STACKNeo4j · Cypher · label-based access · bitmap indexing · pytest (250 tests)
🛡️
ADO → GitHub Secure Migration
Secret-Redaction Pipeline · KLA

An automated migration pipeline that moved a 75-page, 84-attachment Azure DevOps wiki to GitHub — converting syntax, redacting secrets, and surfacing critical credential leaks that were live in the source content.

  • Discovered & remediated 12 critical credential exposures — production passwords, JWT keys and DB credentials.
  • Automated syntax conversion + secret redaction so the migration was safe and repeatable.
  • Processed 75 pages and 84 attachments end-to-end.
12leaks remediated
75pages migrated
84attachments
SecurityAutomationSecret ScanningCI/CD
STACKPython · regex/secret detection · GitHub API · Azure DevOps API · Markdown
🧩
Claude Code Skills Suite
4 Published AI Skills · KLA Internal Registry

Four production Claude Code skills published to KLA's internal AI registry, embedding compliance and best practices directly into the developer workflow — from first install to code review.

  • citizen-dev-review — 33 heuristics across 7 categories with a PLC quality gate.
  • citizen-dev-setup — compliance-from-first-install across 6 JFrog registries.
  • sage-onboarding (15+ systems) & wiki-setup (CI/CD) round out the suite.
4skills shipped
33review heuristics
15+systems onboarded
Developer ToolingComplianceClaude Code
STACKClaude Code SDK · JFrog · CI/CD · policy-as-code · Python
🌈
Rec-Rainbow
LLM Zero-Shot Personalization · Amazon

A session-aware, sequential recommendation agent powered by LLMs — built with a 15-person cross-functional team across the Ads & Prime Video orgs to deliver zero-shot personalization in real time.

  • Zero-shot personalization adapts to a user's live session without per-user retraining.
  • Structured as a data-driven Ads-monetization strategy — every lift tied to revenue.
  • Led a 15-person cross-functional team of scientists & MLEs to production.
+44%CTR
+53%dwell time
+16%revenue
LLMRecSysPersonalizationAds
STACKLLM agents · sequential/session modeling · PyTorch · AWS · A/B testing
💎
HELIUM-PRISM
Multimodal Ranking & Dedup · Amazon (Patented)

A lightweight multimodal model accelerating product mapping, ranking and catalog de-duplication across Kindle & Prime Video for 84 external competitors on 3 continents — outperforming every CNN baseline on both accuracy and efficiency.

  • Patent filed. 98% precision / 99% recall on unmatched decisions, 99% precision on matched.
  • Beat VGG, ResNet, Inception, MobileNet & EfficientNet at 9× speed and 14.5× less memory.
  • Powered de-duplication of a digital catalog across three continents.
98%/99%precision/recall
faster
14.5×less memory
MultimodalPatentVisionEfficiency
STACKMultimodal fusion · CNN backbones · quantization · PyTorch · EMR
📚
AWS Docs RAG & Sherlock
Enterprise Retrieval & Product Matching · Amazon

An enterprise RAG system for AWS Documentation plus Sherlock, a hybrid BERT + ELECTRA product-matching model — together streamlining real-time query answering and slashing catalog onboarding time.

  • RAG over AWS docs (Pinecone + Longformer + Flan-T5) reached 87% BertScore & 0.91 MRR.
  • Sherlock's few-shot matching cut onboarding across 6 content types from 6 months to 3 weeks.
  • A companion BART tag-generation model hit 93% precision@10 & 0.85 MRR.
87%BertScore
0.91MRR
6mo→3wkonboarding
RAGBERT+ELECTRAPineconeNLP
STACKPinecone · Longformer · Flan-T5 · BERT · ELECTRA · BART · few-shot
🧠
Amazon Pay Agent Ecosystem
Multi-Agent Customer Automation · Amazon

A production multi-agent system automating customer queries for 70M Amazon Pay customers — combining intent classification, chat, code-generation and function-calling models, and accelerating multi-hop Bedrock workflows 3×.

  • Zephyr for intent · Mixtral-8x7B chatbot · WizardCoder for code-gen · NexusRaven-V2 for function calling.
  • Anthropic MCP + LangGraph + CrewAI accelerated multi-hop Bedrock workflows and cut OpEx .
  • An AWS Lex integration lifted query-resolution speed 31%, automating 38K support tickets per quarter globally.
70Mcustomers
faster + lower OpEx
38Ktickets/qtr automated
Multi-AgentBedrockMCPFunction Calling
STACKZephyr · Mixtral-8x7B · WizardCoder · NexusRaven-V2 · MCP · LangGraph · CrewAI · AWS Lex
04 — Toolbox

A full-stack AI arsenal.

From the math and the model to MLOps, infra and the boardroom.

🧠 Frontier & Reasoning LLMs

GPT-5Claude 3.5/4Gemini 2.0 Flasho1 / o4DeepSeek-R1/V3Llama 4Grok-3AWS NovaCommand R+Mixtral 8x22B

🪶 Small & Open-Weight

Phi-4Gemma 3Qwen 2.5/3Mistral Large 3Falcon 3Nemotron-4JambaMoE

🕸️ Agentic Frameworks

LangGraphLangChainCrewAIAutoGenAutoGPTSemantic KernelSmolagentsPhidataMCP

🎛️ Fine-Tuning & Alignment

SFT / IFT / PEFTLoRA / QLoRA / DoRARLHF (PPO)DPO / KTO / ORPORLAIF / RLVRKnowledge Distillation

🔎 RAG & Prompting

RAGGraphRAGChain-of-ThoughtTree-of-ThoughtFew-shotReActSemantic Re-ranking

🛡️ Eval & Safety

HELMMMLUTruthfulQARagasRed-teamingGuardrailsExplainable AI

📊 Machine Learning

ClassificationRegressionRandom ForestXGBoostLightGBMCatBoostSVMA/B TestingMulti-Armed BanditCausal Inference

🔬 Deep Learning

CNNLSTMGANMamba (SSM)CLIP / SigLIPLatent DiffusionFlow MatchingGNN / GATWhisper-v3FSDP / DeepSpeed

💬 NLP & Transformers

BERTRoBERTaBARTT5XLNetELECTRAWord2VecfastText

👁️ Computer Vision

YOLOv3Faster/Mask R-CNNUNetEfficientDetViTResNet / EfficientNetSLAM3D Reconstruction

🎯 RecSys & RL

NeuMFBERT4RecRankingReinforcement LearningDeep RLNDCG / mAP / MRR

🐍 Libraries & Frameworks

PyTorchTensorFlowJAXHuggingFaceLlamaIndexvLLMUnslothscikit-learnPolarsW&B

☁️ Cloud & AI Platforms

AWS BedrockSageMakerInferentia2 / TrainiumGCP VertexAITPUDatabricksSnowflake

AI Infra & Serving

CUDAAWQ / GPTQFlash / Paged AttentionGQA / MQAONNXTensorRTTritonvLLMOllama

🔧 MLOps

MLflowAirflowRayKubeflowBentoMLTorchServeFastAPILangSmith

🗄️ Data & Vector Stores

QdrantWeaviatePineconeChromaDBElasticsearchNeo4jPostgreSQLRedis

🌊 Distributed Systems

SparkFlinkKafkaDelta LakeIcebergTrino / PrestoHadoopHorovod

🖥️ Languages & DevOps

PythonC++RustGoTypeScriptScalaDockerKubernetesTerraformGrafana

🎯 Strategy

Corporate StrategyProduct RoadmappingGo-to-MarketTAM/SAM/SOMCompetitive AnalysisM&A Due Diligence

💰 Business & Finance

P&L ManagementFinancial Modeling (DCF/LBO)SaaS Unit EconomicsPricingROI AnalysisOKR/KPI

👥 Leadership

Team BuildingMentorship (45+)Change ManagementExecutive StorytellingC-Suite PresentationsConflict Resolution

📦 Product

AI/LLM ProductizationPLMAgile / ScrumJira / ConfluenceData GovernanceUX & Design Thinking

🛡️ Risk & Compliance

GDPRCCPAHIPAARisk AssessmentSecurityIP Management

⚙️ Operational Excellence

Lean Six SigmaKaizenRCAScenario PlanningVendor ManagementCapacity Planning
My daily-driver stack
🐍Python 🔥PyTorch 🤗HuggingFace 🕸️LangGraph 🚢CrewAI vLLM 📚LlamaIndex ☁️AWS Bedrock 🧊SageMaker 🔗Neo4j 🌲Pinecone 🐳Docker 🐍Python 🔥PyTorch 🤗HuggingFace 🕸️LangGraph 🚢CrewAI vLLM 📚LlamaIndex ☁️AWS Bedrock 🧊SageMaker 🔗Neo4j 🌲Pinecone 🐳Docker
☸️Kubernetes 🌊Spark 📨Kafka 🧱Databricks ❄️Snowflake 🚀Triton 🧠Claude 💎Gemma 🌀JAX 📊MLflow 🌬️Airflow 🦀Rust ☸️Kubernetes 🌊Spark 📨Kafka 🧱Databricks ❄️Snowflake 🚀Triton 🧠Claude 💎Gemma 🌀JAX 📊MLflow 🌬️Airflow 🦀Rust
05 — Writing

Thinking out loud on where AI is going.

I write about the systems-level realities of agentic AI, LLM infrastructure and the future of GenAI products.

06 — Recognition

A track record of being #1.

🏆
Amazon Employee of the Quarter ×4
Across Prime Video, Music & Kindle orgs, globally — 2024.
🥇
Rank 1 of 1,000,000 — 12th Board
99.2%, State Rank 1, West Bengal. Governor's Swami Vivekananda Rashtriya Purashkar.
🎓
CS Department Rank 1 — BITS Pilani
Top of the Computer Science department across ME & BE.
🚀
Hackathon Champion — Amazon '22
1st Prize (of 300+) for "Auto Generation of Video from Text" + Think Big Award.
🌟
Star Performer · WOW · Budding Star
Org-wide performance recognitions, 2021–2023.
🥈
2nd Prize, ProdCon India
National product-pitch competition for a Think Big prototype — Nov 2022.
🔭
Rank 4 of 1,000,000 — 10th Board
96.7%, State Rank 4, West Bengal.
🧪
KVPY Scholar · Top-1% Olympiads
National Astronomy, Physics, Chemistry & Math olympiads.
07 — Education

Foundations in business & engineering.

STEM MBA
The Ohio State University — Fisher College of Business
Finance · Product Management · Strategy · Aug 2025 – Apr 2027 · Columbus, Ohio
Fisher Leadership FellowGraduate AssistantGPA 3.9/4GRE 336GMAT-FE 685
M.E. & B.E., Computer Science
BITS Pilani
Rajasthan, India · Dual degree
CS Dept Rank 1CGPA 9.3/10
08 — Let's build

Let's create something that endures.

Whether it's an autonomous agent, a GenAI product, or an AI strategy that moves the P&L — I'd love to talk.