Kartikeya Agarwal

Applied Machine Learning Engineer with 3+ years of experience architecting large-scale production ML systems, GenAI pipelines, and predictive risk engines at Navi Technologies. Engineered an automated 10-stage Databricks LLM evaluation DAG achieving 87.7% template accuracy and 99.0% precision on Tier 1 financial entities (-74.3% error reduction).

Formulated a Cobb-Douglas template prioritization function and greedy submodular selection algorithm, expanding weekly intake ~20x (from 5k to 105k+ templates) and boosting SMS population coverage from 4.5% to 91.3% across 626M weekly messages. Architected a Flat-Array Compressed Sparse Row (CSR) Trie engine in Scala scaling across 8.39M templates on 100-node Databricks Spark clusters, speeding up feature generation by 5.08x (-80.3% latency). Drove ₹180 Cr ($21M+) in annualized loan disbursal uplift, improved credit approval rates by 150 bps, and saved $550,000+ in annual AWS compute costs.

0+

Years Experience

0M+

Weekly Messages

$0K+

Yearly AWS Savings

0bps

Approval Uplift

Python Scala Kotlin SQL PyTorch Transformers QLoRA Databricks DAGs Multi-Agent LLMs Submodular Optimization CSR Trie Apache Spark (100-node) Delta Lake Kafka Ray Distributed Tensors LightGBM & XGBoost FastAPI MLflow AWS Docker PostgreSQL

Navi Technologies

Machine Learning Engineer 2 · July 2023 – Present · Bengaluru, India

Featured GenAI & Agentic Systems

Automated LLM Annotation DAGs & Multi-Agent Underwriting

Databricks 10-Task DAGgpt-5.6-terradeepseek-v4.1-flashFew-Shot (k=2)Chain-of-ThoughtMulti-Agent

Automated LLM Annotation & Verification DAG (Terra & JEV)

Replaced manual human annotation queues (~700/day, 20.1% baseline accuracy) with an asynchronous 10-task Databricks DAG orchestrating zero-shot hierarchical classification (gpt-5.6-terra), category-scoped few-shot entity extraction ($k=2$), dual-pass fast discriminator review (deepseek-v4.1-flash), zero-cost deterministic rule engines, and reasoned chain-of-thought verification on modified templates. Surged template accuracy to 87.7% (90.0% on core tags), achieved 99.0% precision and 98.5% recall on Tier 1 financial entities (amount, balance, accountNo), and reduced tag errors by 74.3% with 0% review stalls.

Multi-Agent LLM Underwriting Journey & Conversational Guidance

Developed an end-to-end guided re-verification platform for rejected loan applicants, coupling an in-house ITR & GST parser (IGNIS) with a multi-agent LLM underwriting engine and multilingual chatbot for personalized document retrieval. Orchestrated a dynamic 3–5 model ensemble generating 150+ features, delivering a +36 bps credit approval rate uplift, ₹180 Cr ($21M+) in annualized loan disbursal uplift, and ₹10–15 Lakhs in monthly operational cost savings.

87.7% Template Accuracy

99.0% Tier 1 Precision

₹180Cr+ Disbursal Uplift

-74.3% Tag Error Reduction

Distributed Systems & Algorithmic Optimization

Submodular Pareto Selection & Flat-Array CSR Trie Engine

Submodular OptimizationCobb-Douglas ModelingScala Flat-Array CSR Trie100-Node Spark ClusterGemma-4-12B

Algorithmic Template Selection & Submodular Optimization

Solved a critical coverage bottleneck across 626M weekly SMS (where 53–69% are net-new templates weekly) previously throttled by rigid heuristic filters; formulated a novel Cobb-Douglas priority production function balancing 5-dimension local Gemma-4-12B rubric scoring with log-dampened population and user reach. Implemented a greedy submodular marginal-gain selection algorithm along the Pareto frontier, expanding weekly intake ~20x (from 5k to 105k+ templates) and boosting SMS population coverage from 4.45% to 91.3% (+85.5% coverage gain) while capturing high-intent loan overdues, disbursements, and credit limits.

High-Performance Compressed Sparse Row (CSR) Trie Engine

Eliminated reflective 𝒪(T × W) linear template scanning across thousands of classes by engineering an 𝒪(W) Flat-Array Compressed Sparse Row (CSR) Trie in Scala using contiguous primitive arrays (childOffset[], edgeLabel[], edgeTarget[], sChild[], nodeTemplate[]) loaded via bulk binary memory mapping. Scaled across 8.39 million templates and 138,721 buckets (237.5M nodes) on a 100-node Databricks Spark cluster (r-fleet.2xlarge workers), cutting feature generation latency by 80.3% (5.08x speedup from 1,586ms to 312ms) with 100% exact parity across 249,689 test samples.

91.3% Population Coverage

626M Weekly SMS Handled

5.08x Feature Speedup

8.39M Templates on 100 Nodes

Credit Risk & Income Predictive Modeling

Advanced Navi's flagship Prism credit risk scorecard by engineering novel temporal features and replacing univariate retraining pipelines with a multi-modal gradient boosting architecture, directly improving credit approval rates by 150 bps. Built an automated income prediction model using XGBoost and LightGBM, increasing prediction accuracy within a 10% tolerance threshold by 6 percentage points.

+150 bps Credit Approval Rate Increase

+6% pts Income Prediction Accuracy

Multi-Modal Gradient Boosting Scorecard

SPADE Real Time Serving

Co-architected “SPADE Real Time,” a 3-tier reactive Kotlin microservice suite serving SMS, app, and location features at sub-20ms P99 latency at 10,000+ RPS over Kafka and PostgreSQL. Built an automated zero-drift feature store with 99.99% dev/prod consistency that ran 50% faster at half the infrastructure cost.

<20ms P99 Latency @ 10K+ RPS

3-Tier Reactive Kotlin Architecture

99.99% Dev/Prod Consistency

Cloud Cost Optimization & Ray Acceleration

Designed a Databricks compute regression monitor and access attribution framework that secured $550,000+ in annualized AWS savings. Optimized feature selection jobs (Boruta, RFE, IV) using Ray and distributed tensors to slash compute time and training costs by 80%.

$550K+ Annualized AWS Savings

80% Compute Time & Cost Cut

Ray Distributed Tensor Optimization

Automated Feature Store & MLOps Platform

Engineered an end-to-end automated MLOps platform and zero-drift feature store with metrics, alerts, and model versioning via FastAPI, MLflow, and PostgreSQL. Automated pipeline promotion between QA and prod, eliminating manual deployment overhead entirely.

Zero Manual Deployment Overhead

50%+ Faster Pipeline Execution

2x Cost Efficiency

01

Auto Code Sequence Generator

Attention-based Transformer model synthesizing production-ready UI code from natural language prompts and wireframe layouts; designed a customized token vector mapper supporting seamless cross-compilation into React and Angular frameworks.

AttentionTransformersCode GenerationReact & AngularNLP

02

Visual Aid for Visually Impaired

Edge embedded IoT system using ESP32 and Arduino-Nano transmitting wireless camera feeds to an inference server running YOLOv3 and Transformer models for real-time scene understanding and auditory field description trained on MS-COCO.

ESP32 & ArduinoEdge IoTYOLOv3TransformersMS-COCO

03

AI Image Caption Bot

Deep multimodal sequence-to-sequence captioning model combining CNN visual feature extractors with recurrent LSTM networks trained on the Flickr30k benchmark (1.5M parameters).

CNN + LSTMSeq2SeqFlickr30kComputer VisionNLP

Indian Journal of Computer Science

Early Detection of COVID-19 using Machine Learning

Utilized a fine-tuned ResNet-50 deep architecture to extract visual radiographic features from chest X-rays, achieving 99% validation accuracy in differentiating COVID-19 pathologies from normal and bacterial pneumonia scans.

May – July 2021 · ResNet-50 Read Paper (DOI)

arXiv Preprint

Classification of Skin Cancer Images using Convolutional Neural Networks

Developed a two-stage deep learning framework leveraging XceptionNet for lesion image segmentation and a custom Dense CNN for binary benign vs. malignant classification, achieving an 86% validation accuracy.

Apr – May 2021 · XceptionNet Read on arXiv

Academic Standing

CGPA: 8.72 / 10.0 — Graduated with Distinction in Computer Science.

CVSPK Merit Scholarship (2020)

Recipient of prestigious merit scholarship granting 100% Tuition Fee Waiver for academic excellence.

I'm always open to discussing ML engineering roles, interesting projects, or opportunities to collaborate. Let's build something remarkable.

kartikeya72001@gmail.com

About

Experience