Karthik Mulugu — AI Engineer · Machine Learning

01 — Now

Current role

AI/ML Engineer

Socotra Tech Services LLC · May 2026 — Present

Multi-agent booking platform on Amazon Bedrock — customers describe a trip in plain English; agents plan, validate and book across the whole business.

Working with

Amazon BedrockStrandsAWS CDKClaudeGitHub ActionsDockerPython

34K

customers

20K

bookings

12×51

domains × tools

02 — Selected Experience

Shipped a multi-agent booking platform on Strands + Amazon Bedrock AgentCore (Claude Haiku with Sonnet fallback) routing across 12 business domains and 51 tools — 34K customers, 20K bookings in natural language.
Built a document OCR service on AWS Textract + Lambda across 4 government ID types — 10K+ onboardings cleared with zero manual review.
Hardened security end-to-end: JWT Bearer passthrough on live APIs, GitHub OIDC for CI/CD, zero static secrets; consolidated 3 CloudFormation stacks into 2 with AWS CDK.
Resolved 5 separate production failures into one fully documented, stable environment.

Amazon BedrockStrandsAWS CDKClaudeGitHub ActionsDockerPython

34K

customers

20K

bookings

12×51

domains × tools

Took JBL Harman's Salesforce Einstein Bot from failing all 40 pre-launch scenarios to zero critical failures at production launch.
Built 5 automation workflows on Kore.ai (OpenAI + Workday APIs) that cut manual HR ticket handling 40% for a 1,200-person workforce at CBE Companies.
Shipped hybrid RAG pipelines with Azure OpenAI + Search AI, lifting retrieval accuracy 35%.
Deployed guardrails + hallucination detection that cut incorrect answers 25% across 8K monthly queries.

RAGAzure OpenAIKore.aiSalesforce EinsteinWorkday API

40%

less manual tickets

+35%

retrieval accuracy

-25%

wrong answers

Cut mortgage document review from 4+ hours to under 15 minutes per batch across 1,000+ financial documents.
Raised OCR accuracy +35% (Tesseract/EasyOCR tuning) and retrieval precision +40% with a LangChain + LlamaIndex + FAISS RAG pipeline.
Benchmarked Gemini vs Mistral on 1,000+ real production queries — shipped the stronger model, lifting answer accuracy 28% and cutting unanswered queries from 22% to 6%.

PythonOCRRAGLlamaIndexFAISSGemini

4h→15m

review time

+40%

retrieval precision

+28%

answer accuracy

03 — Research

IEEE Access · Under Review

DocuQuery: Hybrid Fusion Retrieval for Document QA

Apr 2026

10.13140/RG.2.2.21211.73766↗Copied

View full paper

Hybrid retrieval with a fixed fusion weight fails silently: lexical-heavy blends starve conceptual queries, dense-heavy blends starve exact ones. DocuQuery fixes this with query-aware fusion gating, where a lightweight classifier routes each query to the optimal mix of BM25, dense, and TF-IDF candidates before cross-encoder reranking, with LangGraph orchestrating multi-intent workflows like summarisation, comparison, refinement, and QA from end to end.

Benchmarked lexical vs. dense retrieval across diverse query types and identified fixed-weight hybrid fusion as a critical failure mode.
Proposed query-aware fusion gating — per-query weighting across BM25, FAISS and TF-IDF before cross-encoder reranking.
75% faster document analysis and +30% retrieval precision over pure dense retrieval on complex multi-page PDFs.

3-way

hybrid fusion

75%

faster analysis

+30%

retrieval precision

RAGLangGraphGeminiFAISSPineconeCross-encoders

04 — Education

University at Buffalo

MS, Computer Science — AI/ML

Graduate coursework in machine learning, deep learning and NLP.

3.7 / 4.0

Jan 2024 — Jun 2025

JNTU Hyderabad

B.Tech, Information Technology

Top 1% in the IT Department batch.

Aug 2019 — Jul 2023

05 — Project Index

Item 01/10Agents

Incident Copilot

Production-ready incident response copilot that unifies logs, metrics and on-call knowledge into one conversational command surface — grounded retrieval, not guesswork.

Unified incident context — logs, metrics, runbooks — into a single conversational surface with grounded retrieval, cutting time-to-context during live incidents.
ChromaDB vector store keeps every answer anchored to your actual infrastructure data, not model memory.
FastAPI backend with streaming responses for real-time triage.

RAGChromaDBFastAPIStreaming

Case study Source

Item 02/10Agents

Contract Risk Reviewer

Multi-agent system that reviews contracts like a junior counsel — clause extraction, risk scoring and plain-English recommendations in minutes, not billable hours.

Multi-agent pipeline handles clause extraction, risk classification and recommendation synthesis end-to-end.
Structured risk reports flag high-stakes clauses with severity scores and suggested redlines.
Turned hours of manual review into minutes with an auditable agent trace.

Multi-agentLLMLegalTechPython

Case study Source

Item 03/10Vision

PerceptAI

Real-time multimodal perception app combining vision models with a live web interface — see what the model sees, instantly.

Real-time inference pipeline streams model output straight into the browser.
Multi-model architecture pairs a vision backbone with a lightweight explanation layer.
Interactive UI makes perception outputs interpretable, not a black box.

Computer VisionReal-timePython

Case study Source

Item 04/10Forecast & ML

Hospital Cost Prediction

End-to-end ML pipeline predicting patient costs on 2.5M real NY state records — SHAP-explained, uncertainty-aware, built for clinical finance decision-support.

Cut RMSE 33% over a Ridge baseline by benchmarking 5 model families with Optuna — LightGBM landed at R²=0.876, RMSE=$7,335 on 2.5M NY SPARCS records.
SHAP TreeExplainer + quantile regression deliver calibrated 80% prediction intervals so clinicians act on uncertainty, not point estimates.
One-command reproducible deployment: FastAPI inference, 6-tab Streamlit dashboard and MLflow tracking in Docker Compose.

LightGBMSHAPOptunaFastAPIMLflow

Case study Source

Item 05/10RAG & NLP

DocuQuery AI Assistant

Production RAG system for natural-language QA over PDFs — hybrid retrieval with LangGraph multi-intent orchestration. Submitted to IEEE Access.

75% faster document analysis via LangGraph multi-intent routing (summarise, compare, refine, QA) backed by Gemini — no manual query-type switching.
Hybrid BM25 + FAISS + TF-IDF with cross-encoder reranking lifted retrieval precision 30% over pure dense retrieval.
Query-aware fusion gating — a novel approach born from benchmarking fixed-weight hybrid failures — submitted to IEEE Access.

LangGraphGeminiFAISSPineconeRAG

Case study Source

Item 06/10RAG & NLP

Sentiment Analysis — DistilBERT + BiLSTM

Transformer vs. recurrent architectures on 204K HuffPost headlines, with a live RSS pipeline surfacing hour-by-hour sentiment from 6 major outlets.

Fine-tuned DistilBERT beat BiLSTM by 4.4 accuracy points — 80.4% accuracy, 0.919 ROC-AUC on 204K headlines.
Live RSS ingestion tracks sentiment across BBC, Reuters, CNN, NYT, Guardian and AP with hour-by-hour trends.
GitHub Actions CI ships Dockerised FastAPI + Streamlit to Streamlit Cloud on every push.

PyTorchNLPFastAPIDocker

Case study Source

06 — Competencies

57 skills · 6 groups

Programming & Data

01

Python
SQL
Git
Docker
Pandas
NumPy
ETL Pipelines
Tableau

Core CS

02

Data Structures & Algorithms
System Design
Object-Oriented Programming
DBMS

Machine Learning & Statistics

03

Classification
Regression
Clustering
Time Series Forecasting
Feature Engineering
Statistical Analysis
Hypothesis Testing
scikit-learn
XGBoost
LightGBM
SHAP

Deep Learning

04

PyTorch
TensorFlow
Keras
CNNs
RNNs / LSTMs
Transformers
Transfer Learning
LoRA / QLoRA Fine-tuning
Data Augmentation

GenAI & LLMs

05

LLMs
RAG
AI Agents
Multi-Agent Orchestration
LangChain
LangGraph
LlamaIndex
Hugging Face Transformers
Prompt Engineering
Embeddings
Vector Databases
Pinecone
FAISS
Guardrails
MCP
LLM Evaluation

Cloud & MLOps

06

AWS Bedrock
SageMaker
AWS Lambda
AWS Textract
AWS CDK
GitHub Actions
CI/CD
MLflow
Databricks

Currently learning

Terraform

Infrastructure as Code

BigQuery

Data Warehouse

07 — Credentials

25

Verified records across 4 disciplines

08 — Beyond Work

01

ISS Volunteer — University at Buffalo

Aug 2024

Volunteered with International Student Services to welcome incoming students, run orientation logistics and support campus events.

02

Hackathon Coordinator — HackFiesta · TechnoMist 2K23

2023 · Hyderabad

Coordinated a college hackathon end-to-end — participant onboarding, judge logistics and live event operations.

KarthikMulugu

01 — Now

AI/ML Engineer

02 — Selected Experience

AI/ML Engineer

AI Engineer

AI Engineer Intern

03 — Research

DocuQuery: Hybrid Fusion Retrieval for Document QA

04 — Education

University at Buffalo

JNTU Hyderabad

05 — Project Index

Incident Copilot

Contract Risk Reviewer

PerceptAI

Hospital Cost Prediction

DocuQuery AI Assistant

Sentiment Analysis — DistilBERT + BiLSTM

06 — Competencies

Programming & Data

Core CS

Machine Learning & Statistics

Deep Learning

GenAI & LLMs

Cloud & MLOps

07 — Credentials

08 — Beyond Work

ISS Volunteer — University at Buffalo

Hackathon Coordinator — HackFiesta · TechnoMist 2K23

Karthik
Mulugu