MLOps engineer focused on applied AI systems, model reliability, evaluation, and deployment workflows.

I like building the parts around models that make them useful in the real world: data pipelines, APIs, monitoring, evals, release gates, and clear failure visibility. Right now I am going deeper into AI/ML engineering through RAG systems, LLM evaluation, and ML infrastructure projects.

Portfolio: https://goparapukethan.github.io/kethan-portfolio/ | Recruiter brief: https://github.com/GoparapukethaN/kethan-portfolio/blob/main/docs/recruiter-brief.md | Proof ledger: https://github.com/GoparapukethaN/kethan-portfolio/blob/main/docs/proof-ledger.md | Project map: https://github.com/GoparapukethaN/kethan-portfolio/blob/main/docs/project-map.md | Enterprise RAG: https://github.com/GoparapukethaN/kethan-portfolio/blob/main/docs/enterprise-rag-reliability-platform.md | Verification note: https://github.com/GoparapukethaN/kethan-portfolio/blob/main/docs/no-key-verification.md

• Reliable RAG and document intelligence systems
• ML evaluation, regression checks, and model release gates
• FastAPI services for ML/AI workloads
• Dockerized local development and deployment workflows
• MLOps patterns: experiment tracking, monitoring, rollback, and reproducibility

Local-first enterprise-style RAG reliability platform for MLOps runbooks and uploaded documents.

• FastAPI backend and Next.js dashboard for ingestion, query, evals, traces, and metrics
• Provider interface with deterministic local mode plus optional OpenAI/Ollama adapters
• Document upload path for Markdown, text, and PDF inputs
• Eval reports with pass/fail, latency, source coverage, cost, citations, and refusal behavior
• Optional providers only receive citation credit for explicit evidence markers
• Secret-extraction requests refuse even when retrieved evidence contains matching terms
• Verified with 29 Python tests, frontend typecheck/build, CLI smoke checks, separate Docker Compose smoke, and rerunnable Playwright dashboard QA
• Enterprise RAG map: https://github.com/GoparapukethaN/kethan-portfolio/blob/main/docs/enterprise-rag-reliability-platform.md
• Demo walkthrough and dashboard screenshot: https://github.com/GoparapukethaN/ai-reliability-lab/blob/main/docs/demo.md#dashboard-demo
• Verification: https://github.com/GoparapukethaN/ai-reliability-lab/blob/main/docs/verification.md
• Proof artifact: https://github.com/GoparapukethaN/ai-reliability-lab/blob/main/docs/proof/verification-2026-05-20.md
• Repo: https://github.com/GoparapukethaN/ai-reliability-lab

Document intelligence and AI evaluation workspace.

• Next.js dashboard and FastAPI backend
• Document ingestion, chunking, retrieval, citations, answer-fact coverage, and release-gate checks
• Paired JSON/Markdown report artifacts with report list/detail endpoints
• Local deterministic retrieval so the demo can run without API keys
• Verified with 21 backend tests, frontend audit/typecheck/build/static export, static demo data checks, tracked static demo browser QA, and separate Docker smoke
• Live static demo: https://goparapukethan.github.io/applied-ai-eval-lab/
• Demo screenshot: https://github.com/GoparapukethaN/applied-ai-eval-lab/blob/main/docs/assets/static-demo-query-eval.png
• Case study: https://github.com/GoparapukethaN/applied-ai-eval-lab/blob/main/docs/case-study.md
• Verification: https://github.com/GoparapukethaN/applied-ai-eval-lab/blob/main/docs/verification.md
• Repo: https://github.com/GoparapukethaN/applied-ai-eval-lab

Retrieval benchmark runner for comparing RAG configuration choices.

• Compares chunking, embedding, dense/BM25/hybrid retrieval, and optional reranking
• Reports hit rate, MRR, context precision, cached query latency, chunk count, and top configuration
• Generates Markdown, JSON, and Pareto plot artifacts for follow-up analysis
• Compares baseline/current benchmark reports with a regression gate for retrieval-quality changes
• Verified with 37 tests, Ruff, a keyless sample benchmark, and a repeatable sample regression gate
• Sample benchmark: https://github.com/GoparapukethaN/rag-forge/blob/main/docs/sample-benchmark.md
• Regression gate: https://github.com/GoparapukethaN/rag-forge/blob/main/docs/sample-regression-gate.md
• Case study: https://github.com/GoparapukethaN/rag-forge/blob/main/docs/case-study.md
• Verification: https://github.com/GoparapukethaN/rag-forge/blob/main/docs/verification.md
• Repo: https://github.com/GoparapukethaN/rag-forge

Local inference serving and benchmark sweeps in plain Python.

• Adaptive batching for WebSocket and HTTP prediction paths
• Backpressure controls for HTTP/WebSocket rate limits and pending-request guards
• Model hot-swap endpoint with service metrics
• JSON load-test reports for local latency and throughput checks
• Deterministic LLM-style benchmark sweeps for batch size and timeout tradeoffs
• JSON and Markdown benchmark reports, with a tracked sample sweep artifact
• Benchmark gate for throughput, p95 latency, error, and recommendation-change checks
• Verified with 40 tests, Ruff, live smoke checks, Docker smoke, sample benchmark gate, and a small load-report run
• Case study: https://github.com/GoparapukethaN/streaminfer/blob/main/docs/case-study.md
• Sample sweep: https://github.com/GoparapukethaN/streaminfer/blob/main/docs/sample-inference-sweep.md
• Sample gate: https://github.com/GoparapukethaN/streaminfer/blob/main/docs/sample-inference-gate.md
• Verification: https://github.com/GoparapukethaN/streaminfer/blob/main/docs/verification.md
• Repo: https://github.com/GoparapukethaN/streaminfer

Pre-deployment release-gate checks for ML models.

• Data drift checks
• Performance regression checks
• Latency regression checks
• CLI-style release gate with Markdown and JSON deployment reports
• Composite action uploads both Markdown and JSON release-gate artifacts
• Missing baseline fails fast unless drift-only mode is explicit
• Local verification covers 26 tests, Ruff, JSON summary counts, CLI/action metadata artifact boundaries, and the sklearn example
• Case study: https://github.com/GoparapukethaN/mlguard/blob/main/docs/case-study.md
• Verification: https://github.com/GoparapukethaN/mlguard/blob/main/docs/verification.md
• Repo: https://github.com/GoparapukethaN/mlguard

Customer churn prediction pipeline with model training, API serving, monitoring, and deployment-oriented project structure.

• Data ingestion and preprocessing
• Model training and experiment tracking
• FastAPI prediction service with request validation and unavailable-model checks
• Prometheus config, Dockerfile, Docker Compose, Kubernetes deployment artifacts, and optional Docker smoke verification
• Local verification covers 15 tests, strict lint/format checks, training import, and Prometheus parsing without requiring Docker
• Optional Docker/Compose checks cover container config, health, and prediction paths
• Case study: https://github.com/GoparapukethaN/mlops-end-to-end-pipeline/blob/main/docs/case-study.md
• Verification: https://github.com/GoparapukethaN/mlops-end-to-end-pipeline/blob/main/docs/verification.md
• Repo: https://github.com/GoparapukethaN/mlops-end-to-end-pipeline

Experimental model-routing CLI and developer-tooling playground.

• Explores provider routing, local tool execution, cost tracking, project memory, and CLI workflows
• Local verification passes Ruff, 5,817 tests, and CLI smoke checks
• Public status doc tracks non-gating mypy, Bandit, format, and live-provider validation gaps
• Showcase status: https://github.com/GoparapukethaN/prism-cli/blob/main/SHOWCASE_STATUS.md
• Repo: https://github.com/GoparapukethaN/prism-cli

Open upstream PRs proposing focused fixes in AI infrastructure and evaluation tooling.

• Ray/RLlib: PR clarifies set_extra_model_outputs behavior against the current implementation
• Hugging Face LightEval: PR fixes an invalid callable type annotation in the parallelism helper
• BentoML: PR adds Python API docs for starting local HTTP/gRPC servers and creating clients
• BentoML: PR adds testing docs for mocking decorated API method bodies
• BentoML: PR adds bentoml.Model API reference coverage for export/import methods
• Ray PR: ray-project/ray#63524
• LightEval PR: huggingface/lighteval#1239
• BentoML server API PR: bentoml/BentoML#5616
• BentoML testing PR: bentoml/BentoML#5617
• BentoML Model API PR: bentoml/BentoML#5618

I am interested in the space between model quality and production reliability. A model can look good in a notebook and still fail once it meets messy data, latency constraints, retrieval misses, unclear evals, and release pressure. My projects are aimed at closing that gap with measurable workflows.

• What I Learned Building Evals Before Adding an LLM
• Enterprise RAG Reliability Platform
• Verifying AI Systems Without API Keys
• AI Reliability Platform Architecture Video Script
• Applied AI Eval Lab Case Study
• Applied AI Eval Lab Demo Script
• RAG Forge Case Study
• StreamInfer Case Study
• MLGuard Case Study
• MLOps Pipeline Case Study

Python, FastAPI, Docker, Kubernetes, SQL, scikit-learn, XGBoost, PyTorch basics, RAG, retrieval evaluation, model monitoring, MLflow-style experiment tracking, GitHub, and cloud-oriented deployment patterns.

• Portfolio: https://goparapukethan.github.io/kethan-portfolio/
• GitHub: https://github.com/GoparapukethaN
• LinkedIn: https://www.linkedin.com/in/kethan-goparapu/