Rohit Patel
Sr. Staff SRE / AIOps Engineer · MLOps & LLMOps · AI Infrastructure · ML Systems · Inference/Serving · Performance · Research · Security
10+ years of SRE/DevOps across Singapore fintech and data-centre infrastructure, now building applied AIOps, MLOps, and AI-security systems end-to-end - backed by original research in RAG security and AIOps root-cause analysis (two arXiv preprints, one first-author).
Open to AIOps, ML Systems, Inference/Serving & AI Security roles - SingaporeConnect on LinkedIn
Full experience history across F5 Networks, Preqin, FundingSocieties, and my current AIOps work at Equinix - plus updates on this research and portfolio.
View LinkedIn Profilegithub.com/rpaut03l
Applied AIOps/MLOps builds, RAG security research, and self-directed AI/ML study repos.
View All RepositoriesRAG-Shield (private research repo)
3-ring defense-in-depth against RAG knowledge-base poisoning - in progress, see the Research tab.
gan-mode-collapse-demo
Interactive live demo simulating GAN mode collapse in real time, with a runnable simple GAN + DCGAN underneath.
ts_02_safevault-reentrancy-demo
Live demonstration of a reentrancy exploit and its defense on Ethereum smart contracts, built and proven end-to-end with Foundry, Anvil, and a MetaMask UI.
mlops-emotion-pipeline
End-to-end MLOps pipeline fine-tuning distilbert-base-uncased for 6-class emotion classification - see the Research tab for the full tech stack and live links.
mlops-assignment2
DistilBERT fine-tuned for Goodreads genre classification - same reproducible MLOps discipline, see the Research tab for links.
rptl_gn_mlops
ML-driven K8s CI/CD pipeline - auto-scaling via Prometheus predict_linear, OTEL instrumentation, versioned model artifacts per PR.
TS-02
Applied deep-learning-ops & MLOps study repo - PyTorch, experiment tracking, distributed training, TorchScript/ONNX export.
TS-01
AI/ML foundations study repo - search, CSP, Bayesian networks, reinforcement learning, convex optimization.
Resume - AI Forward Deployed Engineer / MLOps / AIOps
Sr. Staff SRE background, applied AIOps/MLOps portfolio, and current RAG-security research. Hosted on Google Drive, always the latest version.
View / Download ResumeApplied research and end-to-end builds spanning RAG security, AIOps root-cause analysis, generative modeling, blockchain security, GPU/inference systems, and production MLOps - published work alongside active research, each grounded in something built and tested, not just studied.
TriShieldRAG: A Three-Ring Defense-in-Depth Framework Against Knowledge Corruption in Retrieval-Augmented Generation
Reproduces the PoisonedRAG knowledge-corruption attack on RAG systems end-to-end, then designs and evaluates a defense an attacker must beat on three independent, black-box checkpoints. Posted to arXiv (cs.CR, cross-listed cs.AI / cs.CL / cs.LG); in preparation for peer-reviewed submission to NDSS and IEEE S&P.
I led the design and implementation of TriShieldRAG's three-ring defense architecture - Rings 1–3 above, the paper's core technical contribution - including the retrieval/consensus engineering and the interactive attack/defense demo. Author order reflects the collaborative course/research context; the defense system itself was my build.
TriCalRAG: A Three-Strategy, Retrieval-Augmented Benchmark for On-Premise LLM-Based Root Cause Analysis in AIOps
Benchmarks open-weight LLMs (Qwen2.5-14B, Mistral-Small) served entirely on-premise via vLLM on a single workstation GPU against a classical LSTM log-anomaly detector, across four real production log datasets (BGL, HDFS, Thunderbird, OpenStack) - showing retrieval-augmented prompting is the most reliable configuration for private, on-prem AIOps root cause analysis.
First author and corresponding author on this one - benchmark design, on-premise vLLM serving setup, and evaluation across all four log datasets were my build.
TriParseGPU: GPU-Native Log Template Mining for Large-Scale System Logs
Problem: standard log-pattern grouping processes one line at a time on a single CPU core. Solution: splits the job into three GPU-friendly stages (sort, similarity graph, connected components) - 200x faster, with equal grouping accuracy across all four LogHub benchmark datasets.
TriScaleRAG: Kubernetes-Native Distributed RAG on llm-d + vLLM
Problem: RAG for root-cause analysis works on one machine, but scaling it across a Kubernetes cluster without latency collapse is unproven. Solution: benchmarks three scaling strategies - separate prefill/decode pools, a tiered KV cache (GPU to CPU to shared storage), and prefix-aware request routing - to find what wins at what scale.
TriLore: How Much Organisational Knowledge Does an AI Agent Need to Fix an Incident?
Problem: AI agents can diagnose production faults, but fixing one needs a company's own runbooks and history, and how much is actually required is unknown. Solution: feeds an agent 0%, 25%, 50%, and 100% of an organisation's internal knowledge and measures repair success on unseen fault types, producing a dose-response curve instead of a yes/no claim.
TriCascadeEdge: Cascaded Generative AI for Edge and On-Device Inference
Problem: edge devices can't run large models, forcing a choice between a weak local model or a slow, costly cloud round trip. Solution: a three-tier cascade - on-device, site, and central - like an L1/L2/L3 cache, measured on accuracy, latency, and energy per tier.
TriSplitServe: Prefill/Decode Disaggregated LLM Serving on a Single GPU
Problem: LLM serving's compute-heavy prefill phase and memory-bound decode phase are usually separated across GPUs, but most teams only have one. Solution: tests three levels of on-GPU isolation between the two phases, injects faults to see what breaks, and reports where single-GPU disaggregation actually pays off.
Mode Collapse in GANs - Interactive Presentation Kit
Solo-built interactive live demo simulating GAN mode collapse in real time, with runnable PyTorch implementations underneath.
SafeVault - Reentrancy Exploit & Defense on Ethereum
Live demonstration of a reentrancy exploit and its defense on Ethereum smart contracts, built and proven end-to-end with Foundry, Anvil, and a MetaMask UI.
View RepoMLOps Emotion Pipeline - DistilBERT Fine-Tuning & CI/CD
End-to-end MLOps pipeline - fine-tuning distilbert-base-uncased on the dair-ai/emotion dataset (6-class text classification), with Kaggle GPU training, Weights & Biases tracking, Hugging Face model hosting, Docker inference, and GitHub Actions CI/CD.
| Layer | Tool |
|---|---|
| Model | distilbert-base-uncased (Hugging Face Transformers) |
| Dataset | dair-ai/emotion - split config (20k) |
| Training | Kaggle Notebook, GPU T4, HF Trainer API |
| Experiment tracking | Weights & Biases |
| Model registry | Hugging Face Hub (public) |
| Packaging | Docker (slim Python 3.11) |
| CI/CD | GitHub Actions |
MLOps Assignment 2 - DistilBERT for Goodreads Genre Classification
DistilBERT fine-tuned for multi-label genre classification on the Goodreads dataset, with the same reproducible MLOps discipline - tracked training, public model artifact, and public dashboard.
Live LinksCertifications
13 active certifications across Kubernetes, observability, IaC, and AI - verified and browsable on Credly.
View All Badges on CredlyBrowse by category
- CKA - Certified Kubernetes Administrator
- CKAD - Certified Kubernetes Application Developer
- KCNA - Kubernetes and Cloud Native Associate
- KCA (Kyverno) - Kyverno Certified Associate
- Cloud Native Associate (CNCF)
- LF Kubernetes Specialist
- PCA - Prometheus Certified Associate
- OTCA - OpenTelemetry Certified Associate
- LF Monitoring & SRE
- HashiCorp Terraform Associate
- AWS AI Practitioner
- LF Linux Essentials
- LF Linux Provisioning
Writing
Long-form technical writing on AIOps, MLOps, and applied AI/ML - deployment guides, workflows, and lessons from production systems.
Read on MediumMore writing
- rohitpatel.in - long-form posts on AWS ML Engineer certification prep, ML workflows, and MLOps practice
- Medium profile - cross-posted and standalone technical writing
Comments