I build multi-model LLM orchestration platforms with eval gates that block production when accuracy drops. 30+ models, 128+ users, 1,690 ground-truth samples. Senior AI Infrastructure & Systems Engineer at Airbnb. 9+ years shipping production AI.
AI Infrastructure & LLM Serving
Own end-to-end architecture of BPI Virtual Analyst at Airbnb - 30+ foundation models (AWS Bedrock, OpenAI, Anthropic Claude, vLLM) behind FacadeDriver. 128+ users across 4 teams. 10K rows per run. 50% cost savings via batch inference.
Eval Gates & RAG Pipelines
23-version LLM eval harness with 1,690 ground-truth samples. LLM-as-Judge gates block production at precision < 0.85. End-to-end RAG with vector + hybrid retrieval, reranking, contextual compression. F1 lifted to 0.654 (project-best).
Streaming, Observability & Certifications
Kafka 4M req/min at Southwest. OpenTelemetry, Loki, Datadog at Airbnb. AWS Solutions Architect Professional, AWS ML Specialty, GCP Data Engineer, Azure Data Scientist. 4 publications, 1 patent.
hover or press a key
Years Experience
LLM Models Orchestrated
Users Across 4 Teams
Ground-Truth Eval Samples
Senior AI Infrastructure & Systems Engineer
Owned end-to-end architecture of BPI Virtual Analyst - a multi-model GenAI orchestration platform abstracting 30+ foundation models (AWS Bedrock, OpenAI, Anthropic Claude, vLLM) behind FacadeDriver, serving 128+ users. Scaled from 600 to 10,000 rows/run (40MB uploads, PII-safe). Architected Presidio PII pipeline (30% faster). Led Redpen label export upgrade targeting 80% runtime reduction.
Python · Streamlit · Flask · Celery · Airflow · Labelbox · Presidio · AWS · OTEL
View details
Senior Software Engineer (Contractor) - Dose Management Platform
Built and maintained the Dose Management System (DMS) - full-stack healthcare portal for medication management. Java/Spring Boot backend, React frontend, deployed on OpenShift OCP across dev/QA/prod environments.
Java · Spring Boot · React · OpenShift · PostgreSQL · GitHub Actions
View details
Senior Software Engineer (Contractor) - Backend & Data Platform
Owned architecture and production operation of Kafka pipelines sustaining 4M req/min for real-time flight tracking. Containerized services with Docker + Kubernetes. Cut on-call MTTR from 45 to 12 minutes (73% reduction). Statistical analysis and regression models on large datasets.
Python · Docker · Kubernetes · AWS · Datadog · Flask
View details
Senior Software Engineer (Contractor) - Backend & Data Science
Owned API service handling 17M pageviews/month at 94% cache efficiency. Cleared 200+ bottlenecks; app 5x faster after refactor. Improved NLP accuracy 86% to 94%. Deployed ML models on AWS SageMaker. Published at SPE ADIPEC 2022 Conference (SPE-210986-MS).
Python · PySpark · Azure Databricks · AWS SageMaker · MLFlow · Flask · Docker · Jenkins
View details
Software Engineer, ERP Analytics & Data Engineering
Built ERP analytics dashboard across 13 business units; boosted client activity by 20%. Automated PIP process - saved 600+ monthly work hours. Architected real-time fraud detection pipeline using Kafka. Best Performer Q3 2018.
Python · Java · Oracle Cloud HCM · Kafka · ELK Stack · AWS · Flask · PostgreSQL
View details
Production architectures I built at Airbnb
My work
Real-world projects showcasing skills across GenAI, ML infrastructure, full-stack development, and cloud platforms.
Added token cost tracking and inference latency telemetry callbacks to LangChain's LLM invocation pipeline. Open-source contribution to the world's most widely used LLM orchestration framework. Empowers production teams to monitor model spend across providers in real time.
Engineered ReliabilityObserver for the LiveKit Agents voice AI framework with 23 unit tests. Monitors WebRTC connection degradation, triggers automated reconnection backoff, and streams reliability telemetry for production real-time voice agents.
Fixed Vertex AI Gemini token counting calculation in the LiteLLM proxy gateway. Correct token counts are mission-critical for cost attribution, rate limiting, and context window management across multi-provider deployments. Merged to main branch.
Architected multi-model GenAI orchestration platform abstracting 30+ foundation models (AWS Bedrock, OpenAI, Anthropic Claude, vLLM) behind a model-agnostic FacadeDriver. Scaled to 10,000 rows/run for 55+ analysts with Presidio PII sanitization and Insight Miner.
Owned Airbnb's data-labeling and evaluation platform UI. Flask, Celery, Labelbox, Redis. Architected robust data pipelines with SQLAlchemy and Alembic for model evaluation workflows. Serves CS Labs, Luxe Labelbox, and AirCover HALO teams.
Owned backend architecture for a full-stack healthcare portal handling radioactive isotope distribution under a strict 10-hour decay SLA. Java/Spring Boot backend, React frontend, deployed on OpenShift OCP with 99.9% uptime across clinical environments.
Open-source multi-signal detector for AI-generated images. Three signals (frequency-domain DCT, wavelet reconstruction error, PRNU noise residual) with temperature calibration and refusal verdicts. Evaluated on GenImage benchmark: BigGAN AUROC 0.94.
Owned ML reusable framework for subsurface applications. Refactored Jupyter to production Python package (5x faster execution). Deployed on AWS SageMaker and Databricks. Published at SPE ADIPEC 2022 (SPE-210986-MS).
Computer Vision and NLP application built to assist visually impaired individuals navigate physical environments in real time. First Place Winner at HackNYU.
Indian Patent Office — Application No. 202541026299
Sai Likhith Kanuparthi, Sai Yasaswini Majety (Inventors)
Architectures for lightweight, parameter-efficient adapters enabling models to transfer knowledge dynamically across disparate domains without catastrophic forgetting. Aligns with frontier modular DL research (Google Modular Deep Learning survey, Microsoft US10235994B2). IPC: G06N 3/08, G06N 3/045, G06N 3/096, G06N 3/084, G06N 20/00.
SPE ADIPEC 2022 (Abu Dhabi) — SPE-210986-MS
Jie Chen, Changqing Yao, Pandu R. Devarakota, Mohamed Sidahmed, Zexuan Dong, Sai Kanuparthi, Jia Wei
Reusable, platform-independent ML framework for energy applications supporting local, Docker, and HPC environments. Co-authored with Shell's top ML researchers (lead author h-index 69, 21,611 citations). Demonstrated cost + turnaround time savings in production.
IEEE Xplore (ICoACT 2023) — Document ID: 11004721
Sai Likhith Kanuparthi, Naresh Kumar
Analyzes NLP architectures (LSTMs, Transformers) applied to online harassment detection and discusses the transition to real-time generative counters. Indexed in IEEE Xplore with DOI.
A Case-Based Study of State Space Models in Health Care: The New Transformers (Ch. 1)
Ria Ghosh, Rajeev Kumar, Sai Likhith Kanuparthi
Traces the development of State Space Models (SSMs) from control theory and signal processing to modern ML applications. Covers SSM applications in predictive maintenance, financial forecasting, and healthcare. Opening chapter of a Cambridge Scholars book.
Harnessing Generative AI to Combat Cyberbullying in Industry (Ch. 7) — ISBN 978-1-0364-5866-9
Sai Likhith Kanuparthi, Ria Ghosh, Meetu Malhotra, Naresh Kumar
Proposes next-generation AI moderation frameworks targeting positive reinforcement, agentic defense networks, and proactive intervention. Published in a 468-page edited volume on Generative AI for online safety.
Harnessing Generative AI to Combat Cyberbullying in Industry (Ch. 4, p. 200)
Sai Likhith Kanuparthi, Naresh Kumar
Reviews the failure modes of classical machine learning and heuristic models in high-dimensional text streams, proving why deep autoencoders and generative moderation are required.
2025 International Conference on Advanced Computing Technologies (ICoACT) — Accepted
Preetish Kakkar, Shiwam Mittal, Sethuraman Ulaganathan, Tamilarasan Kannadasan, Sai Likhith Kanuparthi
Synthesizes the architecture requirements for digital twin synchronization, Edge AI data routing, and VR interface streaming under Metaverse frameworks.
Phytomedicine (Elsevier) — Under Review
Jaswanth Krishna Boddupalli, Vijaya Bhaskar Reddy Pichili, Sai Likhith Kanuparthi, Kiranmai Chadipiralla
Employs PCA, machine learning categorization, and metabolic profiling to isolate active compounds in JA/SA-elicited plant suspension cultures. Submitted to Elsevier's top phytotherapy journal (IF ~6.0).
AWS Certified Solutions Architect — Professional
Amazon Web Services
AWS Certified Developer — Associate
Amazon Web Services
AWS Certified Machine Learning — Specialty
Amazon Web Services
Microsoft Certified: Azure Data Scientist Associate (DP-100)
Microsoft
MS in Computer Science
New York University
New York, NY · GPA 3.69 / 4.0
View details
B.Tech in Electronics & Communication Engineering
Jawaharlal Nehru Technological University
Andhra Pradesh, India · GPA 4.0 / 4.0
View details
Contact
I'm currently open to new opportunities. Whether you have a question, a project idea, or just want to say hi — my inbox is always open.
Email directly
sailikhithcse@gmail.comLocation
San Francisco, CA
