Atharva Hirulkar

About me

I build the systems that make ML work in production.

I'm an ML Systems Engineer with hands-on experience building production data infrastructure at serious scale. At State Street Corporation, I owned Snowflake pipelines processing ~5M records/day, monitored ~$1B in real-time transaction flows, and engineered ISO 20022-compliant payment workflows.

I've architected multi-cloud infrastructure across AWS, Azure, and OCI using Terraform - and automated operations with Ansible, cutting provisioning time by ~40%. I care about reliability, not just velocity.

Currently pursuing an MS in Data Science at UC San Diego (GPA 3.80), deepening expertise in scalable data systems, statistical NLP, and optimization. I hold a granted copyright in biomedical time-series ML.

Data Systems

ETLSnowflakeKafkaSparkAirflowPostgreSQLTimescaleDBNeo4jQdrantData Warehousing

Machine Learning

PyTorchTensorFlowScikit-learnMLflowMLOpsNLPTransformersHugging Face LLMsRAGPrompt EngineeringOllamaOpenAI APILangChainAgentic AIFeature EngineeringAnomaly DetectionTime-Series Forecasting

Cloud & DevOps

AWSAzureOracle CloudTerraformAnsibleDockerKubernetesCI/CDGrafana

Programming & Tools

PythonSQLShell / BashPandasNumPyNetworkXPlotlyMatplotlib

Career

Where I've worked
& what I shipped.

Jul 2023 - Aug 2025

State Street Corp.

Bangalore, India

Data Engineer

Built and owned SQL + Python ETL pipelines on Snowflake processing 5M+ records/day, ensuring 99.9%+ uptime compliance with enterprise SLAs across production data workflows.
Architected and provisioned multi-cloud data infrastructure (OCI, AWS, Azure) using Terraform IaC, enabling scalable, zero-downtime deployments across production data pipelines.
Automated data environment provisioning and configuration using Ansible playbooks and Shell scripting, reducing manual operational overhead by 40% across production data environments.
Implemented Azure DevOps CI/CD for data services, cutting release cycles by 40% (5 days → 3 days) and achieving 99.8% deployment success across cloud data workflows.

PythonSQLSnowflakeTerraformAnsibleAzure DevOpsAWS

Jan 2023 - Jul 2023

State Street Corp.

Hyderabad, India

Payment Systems Analyst - Intern

Built real-time monitoring dashboards for $1B daily transaction flows (LYNX, CHIPS, TARGET2), detecting 98% of pipeline failures within 30 seconds and sustaining 99.9% settlement uptime.
Engineered ISO 20022 payment data workflows in Shell and PL/SQL, raising data-validation accuracy from 96% to 99.4% on high-volume financial messaging pipelines.

PL/SQLShellISO 20022Financial Messaging

Work

Things I've built.

SignalStack

End-to-end ML systems pipeline. Polygon.io WebSocket → Kafka → PySpark Structured Streaming → TimescaleDB feature store → three concurrent models (LSTM, LightGBM, Isolation Forest) → FastAPI at <10ms p99. Point-in-time correct features, PSI drift monitoring, live Grafana dashboards.

KafkaPySparkTimescaleDBGrafanaPolygon.ioPython

↗

FraudLens

Production ML system for real-time fraud detection with explainable AI. IEEE-CIS · XGBoost · LightGBM · MLflow · FastAPI · AWS ECS Fargate + ALB · Qdrant · Airflow · GitHub Actions CI/CD.

XGBoostLightGBMFastAPIAWS ECS FargateMLflowAirflow

↗

CadburyIn Development

Local-first AI agent with explicit constraint architecture. Deny-by-default tool policy, allowlist-only file search with user approval, semantic RAG over local folders with source citations, and full audit logging of every tool call. Ollama local inference keeps all prompts and file text on-device. The policy layer is the product.

PythonOllamaQwen 2.5 7BRAGSemantic SearchPolicy Engine

↗

JobPilotIn Development

Autonomous job application pipeline powered by GPT-4o-mini. Resume to JD alignment scoring with structured skill-gap extraction, LaTeX resume tailoring via brace-match injection, Cover letter generation, and PostgreSQL persistence with dedup pipeline.

PythonGPT-4o-miniFastAPIPostgreSQLLaTeXOpenAI API

↗

CosmeTik

Multi-database skincare analytics platform. Unified product + review pipeline across PostgreSQL, Neo4j (entity graphs), and Qdrant (vector search). 1M+ reviews, hybrid ML recommender.

PostgreSQLNeo4jQdrantPythonNLPRecommendation Engine

↗

PulseMLCopyright L-114951/2022 (Gov. of India)

Wearable ECG/PPG physiological monitoring. LSTM-based arrhythmia detection with custom IoT data pipeline. Core ML framework for real-time anomaly alerts on biosignals.

LSTMPyTorchSignal ProcessingIoTPythonFeature Engineering

↗

Multilingual Speech Transcription

Video-to-text pipeline supporting 100+ languages. BART-Large-CNN abstractive summarization with multilingual encoder. Flask API serving transcription + translation.

BARTHugging FaceNLPFlaskPythonTransformers

↗

Seismic Risk AtlasDataHacks 2026

🏆 Best Use of Marimo & Sphinx

Block-level earthquake loss estimator for LA County's 2,498 census tracts. Physics-based ground-motion simulations (500 M6.7 scenarios, Scripps Institution) fused with Zillow ZHVI housing values and ACS census demographics via KD-tree spatial joins on Databricks Spark. FEMA HAZUS fragility curves applied per building code era; Monte Carlo aggregation over all scenarios preserves damage-function nonlinearity. XGBoost damage model GPU-trained on NVIDIA L40s (R² 0.99996, 0.70s). Interactive Leaflet choropleth map with FastAPI + OpenAI tract-level plain-English summaries.

PythonMarimoDatabricksSparkFastAPILeaflet.jsXGBoostSphinxFeature EngineeringOpenAI

↗

Building data systems that actually scale.

I build the systems that make ML work in production.

Where I've worked
& what I shipped.

Things I've built.

Credentials &
in progress.

Curious who I really am?

Got an
interesting
problem?

Building data systems that actually scale.

I build the systems that make ML work in production.

Where I've worked& what I shipped.

Things I've built.

Credentials &in progress.

Curious who I really am?

Got aninterestingproblem?

Where I've worked
& what I shipped.

Credentials &
in progress.

Got an
interesting
problem?