ASK.

Full Stack AI Engineer / Systems Builder

AI is powerful.
Making it reliable
is engineering.

I'm Abhinaysai. I build intelligent products—from agent architecture and context engineering to cloud infrastructure and the evaluation loops that make systems dependable.

Explore selected workStart a conversation
Agentic AI & LLMsCloud & platformFull-stack products

Systems that

Useful action.

plan • act • coordinate

Select a layer to inspect how reliability is built

Reliable Agents

Failure benches, judges, and recovery so agents degrade into diagnosis—not silence.

Context Engineering

MCP memory with ingest, retrieval, provenance, and access control for agent Tools.

AI Infrastructure

Training, serving, and observability that hold under real traffic.

Featured work.

01

Meridian

10K+ suppliers. 250+ risk edges. 500–700 req/s

Real-time supply chain risk platform combining geopolitical signals (GDELT, ACLED), AIS vessel tracking, and weather into a live Neo4j knowledge graph with explainable XGBoost scoring.

EngineeringProductResearchKafkaNeo4jXGBoostFastAPI
10K+ supplier nodesXGBoost + SHAP scoring14 FastAPI endpoints
2026→
GitHub ↗
02

Cortex

10K+ nodes. 30% recall lift. MCP native

MCP-based organizational memory that ingests Slack, GitHub, and Jira through Kafka into a Neo4j knowledge graph and exposes persistent context to agents via a TypeScript MCP server.

ProductEngineeringResearchMCPKafkaNeo4jFastAPI
MCP agent memory10K+ graph nodes30% recall vs baseline
2026→
GitHub ↗
03

EEG Seizure Detection

15+ architectures. AUROC 0.740. 916 hours

Benchmark of 15+ neural architectures on 916 hours of pediatric EEG (CHB-MIT, 24 patients) for seizure detection under strict subject-independent evaluation. Best AUROC 0.740.

ResearchEngineeringDeep LearningPyTorchNeural Signal Processing
15+ neural architecturesBest AUROC 0.740Patient-disjoint evaluation
2026→
GitHub ↗
04

Tool-semantics

Schema and behavior. MCP Tool contracts

Behavioral compatibility testing for MCP Tools and AI agent interfaces. Know when a change breaks the agent, not just the schema.

EngineeringResearchProductMCPEvaluationTool CallingTypeScript
MCP Tool interfacesBehavioral compatibilityAgent breaking change detection
2026→
GitHub ↗

Tools of the trade.

Core first. What I actually use.

Generative AI & Agents
LLMsRAGMCPTool CallingLangGraphEmbeddingsVector SearchPrompt EngineeringLLM Evaluation
ML & Data Science
PyTorchscikit-learnXGBoostFeature EngineeringSHAPMLflowTensorFlowTime-Series
Data Eng & APIs
KafkaAirflowSpark StreamingFastAPINeo4jQdrantBigQueryPostgreSQL
Cloud & MLOps
DockerKubernetesTerraformJenkinsGitHub ActionsAWS BedrockAzure MLGCP Vertex AIVault
Observability
PrometheusGrafanaOpenTelemetryJaegerSLA Monitoring
Languages
PythonSQLTypeScriptBash

Full-stack AI with judgment.

I build reliable AI systems end to end: agents that recover under failure, evaluation that catches silent regressions, and infrastructure that keeps research honest in production.

The work spans product enhancements and applied research on modern AI/ML stacks. Meridian for supply chain risk intelligence. Cortex for organizational memory. Clinical EEG benchmarking. Tool-semantics for MCP interface contracts.

Experience backs the idea. At Follett I ran enterprise fulfillment operations at scale. At GWU I ran graduate research on climate driven species interactions with AI assisted NLP. At Jio I shipped multi cloud MLOps. At Tetra Pak I worked as an Information Technology Engineer on Azure-backed predictive maintenance and data workflows.

85%

Faster model deploys

100+

Airflow DAGs

50M+

Events per day

45%

Lower MTTD

Engineering

Build what holds.

Memory, Tools, evaluation, and infrastructure. The layers that make agent behavior reproducible under real constraints.

Product

Ship the consequence.

Research that ends in a system someone can run, inspect, and depend on. Not a notebook that dies in a week.

Research

Measure what breaks.

Failure benches, behavioral contracts, and observability that catch silent regressions before users do.

Where I’ve shipped.

May 2025 to May 2026

Washington, DC

System Operations Analyst

Follett Higher Education

  • Administered enterprise fulfillment infrastructure, managing over 150K daily transactions while ensuring PCI compliance.
  • Architected automated reporting workflows and real-time dashboards, reducing operational toil by 15 hours weekly.
  • Optimized backend operations through CourseTracks integration, achieving a 25% reduction in inventory exposure.

Mar 2025 to Aug 2025

Washington, DC

Graduate Research Assistant

The George Washington University

  • Analyzed over 5,000 ecological and species-interaction data points across 50+ global ecoregions to model species responses to climate change.
  • Developed AI-assisted Python workflows utilizing NLP and machine learning, reducing manual research processing time by approximately 40%.
  • Supported a large-scale meta-analysis on invasive species interactions, contributing to the understanding of ecosystem-level effects.

Jul 2023 to Jul 2024

Navi Mumbai, India

Data Ops Engineer

Jio Platforms Limited

  • Engineered CI/CD automation for ML model deployment across AWS, Azure, and GCP using Jenkins, Docker, and Kubernetes; cut time-to-production by 85% and release failures by 60%.
  • Deployed Kubernetes-based ML inference microservices for real-time and batch workloads; reduced model-serving downtime 40% with auto-scaling under high-velocity traffic.
  • Built and maintained 100+ Apache Airflow DAGs with Vault-secured authentication for ML and data pipelines, achieving 94% pipeline reliability across multi-terabyte daily volume.
  • Optimized Spark Streaming jobs for real-time feature pipelines, cutting data latency 40% under peak workloads.
  • Provisioned cloud infrastructure via Terraform across AWS and Azure, achieving 30% cost reduction through right-sizing.
  • Integrated Prometheus and Grafana observability with automated alerting, reducing MTTD by 45%.

May 2022 to Jun 2023

Pune, India

Data Analyst

PHN Technology Pvt Ltd

  • Engineered data pipelines in BigQuery reducing ML feature-serving latency by 40%, enabling faster model iteration and real-time inference on production workloads.
  • Designed and secured Apache Airflow infrastructure with key-based authentication, protecting 75% of critical data and ML pipelines while maintaining CI/CD deployment velocity.
  • Built 10+ real-time monitoring dashboards surfacing data quality issues and model drift signals before degradation reached users.

Jun 2021 to May 2022

Pune, India

Information Technology Engineer

Tetra Pak

  • Conducted experiments with supervised learning models, including Random Forest and Gradient Boosting, on sensor data for predictive maintenance.
  • Designed and deployed scalable cloud infrastructure on Microsoft Azure to support enterprise-grade data and AI workloads.
  • Automated data aggregation and reporting workflows using Python and SQL, reducing manual effort by 40%.

Education.

2024 to 2026

Washington, DC

M.S. Data Science

The George Washington University

  • Global Leaders Award. 3.77 GPA.
  • Deep learning and time-series modeling: LSTM, Mamba, MoE, Transformers.
  • Ship work: 15+ architecture benchmarks on 916 hours of pediatric EEG; biomedical imaging detection pipelines.

2019 to 2023

Pune, India

B.E. Computer Engineering with Honors in Data Science

Savitribai Phule Pune University

Contact.

Technical discussions on agent systems, evaluation, or AI infrastructure—email is the fastest path.

EmailLinkedInGitHubXMedium

Report an issue

Found something broken?

Layout bug, dead link, typo, or a suggestion: send it over. I read every one. No mail client? Copy the draft and open a GitHub issue.