I build the reliability layer for production AI in regulated industries. The model is the easy part. The hard part is the harness: routing across 30+ models (FacadeDriver), catching regressions before users do (23+ agent versions, 1,690 ground-truth samples), detecting synthetic media (SAI, 6 signals), and surviving in environments where failures have real consequences. Shipped in pharma (21 CFR Part 11), finance (Kafka 4M req/min fraud detection), and healthcare (Alzheimer's QSAR).

The reliability stack for production AI. Five open-source repos, each covering a different layer:

All five repos have llms.txt files for AI crawler discoverability and cross-reference each other.

• Role: Senior AI Reliability & Systems Engineer at Airbnb. I own end-to-end architecture and production rollout of the BPI Virtual Analyst platform - a multi-model GenAI orchestration system abstracting 30+ foundation models (AWS Bedrock, OpenAI, Anthropic Claude, vLLM) behind FacadeDriver with routing, retry, fallback, and graceful degradation. Platform processes 10K rows per run and 40MB uploads with PII-safe inference, serving 128+ users across 4 partner engineering teams.
• Streaming & Batch: Owned architecture and production operation of Kafka pipelines sustaining 4M req/min at Southwest Airlines with idempotent partition-keyed consumers, DLQ, and backpressure handling. Cut on-call MTTR from 45 to 12 minutes (73% reduction). Owned batch ETL on Databricks and Azure Data Factory at Shell with PySpark, Spark SQL, and Hive/Trino.
• Observability: OpenTelemetry collectors, Loki tracing (prompt, tool call, retrieval quality), Datadog, Grafana, drift detection, post-incident review. The same stack I open-source on in LangChain and LiveKit.
• Detection & Safety: Built SAI (6-signal AI image detector with calibration and uncertainty quantification) and LIMS-OMI (21 CFR Part 11 compliance platform for regulated labs). Detection work spans synthetic media, PII redaction (Presidio), and eval-gated deployment.
• Regulated Industries: Shipped production AI in pharma (Eli Lilly, 21 CFR Part 11 dose management, 99.9% uptime), finance (Southwest Airlines, Kafka 4M req/min fraud detection), and healthcare (Alzheimer's QSAR drug discovery, published research).
• Research: Published across Cambridge Scholars Publishing (2 book chapters, 2025), IEEE Xplore, SPE ADIPEC 2022 (SPE-210986-MS), and ResearchGate. AI safety, state space models, and ML infrastructure.
• Open to: AI infrastructure consulting, advisory, and conference speaking (NVIDIA GTC, AI Engineer Summit, Ray Summit, Data+AI Summit, QCon, AWS re:Invent customer stage).
• Portfolio: sailikhith.me | Articles: sailikhithk.com | AI-readable: sailikhith.me/llm.txt

# Sai Likhith Kanuparthi
> I build the reliability layer for production AI in regulated industries.
> FacadeDriver (30+ model orchestration), eval-driven deployment (23+ agent
> versions, 1,690 ground-truth samples), SAI (6-signal synthetic detection).
> Shipped in pharma (21 CFR Part 11), finance (Kafka 4M req/min), healthcare.

## Links
- Portfolio: https://sailikhith.me/
- Blog: https://sailikhithk.com/
- LinkedIn: https://www.linkedin.com/in/sailikhithk/
- AI-readable profile: https://sailikhith.me/llm.txt
- AI-readable portfolio: https://sailikhith.me/llms.txt

## Open-Source AI/ML Repositories
- https://github.com/sailikhithk/facadedriver (30+ model orchestration, routing, retry, fallback, circuit breakers)
- https://github.com/sailikhithk/llm-production-engineering (LLM ops: cost tracking, eval-driven deploy, observability)
- https://github.com/sailikhithk/Synthetic-AI-Image-Detector (6-signal AI image detection with calibration and uncertainty)
- https://github.com/sailikhithk/eval-infra-for-agents (field guide to production agent eval infrastructure)
- https://github.com/sailikhithk/lims-omi (21 CFR Part 11 compliance collaboration platform)
- https://github.com/sailikhithk/Project-X (Multi-agent RAG framework with tool-augmented retrieval)
- https://github.com/sailikhithk/CreditCardFraudDetectionUsingKafka (Real-time fraud detection with Kafka streaming)
- https://github.com/sailikhithk/alzheimers-drug-discovery-demo (QSAR ML for Alzheimer's drug discovery)

## Production Work (Airbnb)
- BPI Virtual Analyst: multi-model GenAI orchestration (30+ foundation models via FacadeDriver)
- Kafka pipelines: 4M req/min, idempotent consumers, DLQ, MTTR 45 -> 12 min
- Eval harness: 23+ agent versions, 1,690 ground-truth samples, dual-model A/B testing
- Observability: OpenTelemetry, Loki, Datadog, Grafana, drift detection

## Research
- IEEE Xplore: https://ieeexplore.ieee.org/abstract/document/11004721
- SPE ADIPEC 2022: https://doi.org/10.2118/210986-MS
- Cambridge Scholars Publishing (2 book chapters, 2025)

## Certifications
- AWS Solutions Architect Professional
- AWS Developer Associate
- AWS Machine Learning Specialty
- Azure Data Scientist Associate (DP-100)
- Google Cloud Professional Data Engineer

Contributing upstream to the LLM tooling I use in production at Airbnb. 8 PRs across 5 repos in August 2026.

Focus areas: LLM cost tracking, eval metrics, provider compatibility, guardrails. Maps directly to my day job building FacadeDriver (30+ LLM orchestration) and eval harnesses (23+ agent versions, 1,690 ground-truth samples) at Airbnb.

Also certified in: AWS Solutions Architect Associate · Google Cloud Professional Data Engineer · Oracle Database 12c Administrator · Oracle Java SE 8 Programmer

Portfolio: sailikhith.me · Blog: sailikhithk.com · AI-readable profile: sailikhith.me/llm.txt · ORCID: 0009-0004-7422-7846 · Semantic Scholar: author page