building-ai-developer-tools
What I'd Do Differently If I Started Pipeshift Today
I am the founder of Pipeshift -- CI/CD migration intelligence for Jenkins-to-GitHub Actions moves.
Writing
Cloud & AI Engineering Blog
Practical patterns, trade-offs, and lessons learned from real systems across OCI, AWS, and Azure — plus DevOps and agentic AI engineering.
fine-tuning-with-synthetic-data
I Fine-Tuned a Model on Synthetic Data. Here Is What the Numbers Actually Said.
The popular claim is that synthetic data closes the gap between a capable base model and a task-specific one -- no labeling budget, no data collection headaches, just generate-and-fine-tune. That claim is half right.
qdrant-out-of-memory
Qdrant Ran Out of Memory During a Client Demo. Here Is the Four-Hour Postmortem.
The system had been running fine for three weeks. Queries returning in under 200ms, recall looking reasonable in our offline evals, nothing alarming in the logs.
teaching-engineers-to-use-llms
The Four-Week Pattern I Use to Get Skeptical Engineers Actually Shipping AI Features
Most engineering teams that bring me in for AI consulting have the same latent problem: one or two people understand how LLMs actually work, the rest have absorbed enough LinkedIn content to be confused, and management has already promised.
llm-judge-retrieval-quality
The Three-Layer Metric Pyramid for RAG Retrieval Evaluation
Most RAG evaluation frameworks I encounter measure the wrong thing with high precision.
autonomous-ai-agents-failure-rate
I Build AI Agents for Clients. Here Is Why I Advise Against Fully Autonomous Ones in Most Production Systems.
Let me be clear about where I am coming from before I say anything critical: I build agents. I am the founder of [Pipeshift](https://mohakdeepsingh.dev/products), which uses a multi-agent LangGraph pipeline under the hood.
llm-api-cost-reduction
I Cut a Client's LLM Bill by $8k/Month With a Routing Classifier. Here Is What the Benchmark Actually Looked Like.
The pitch for model routing is simple: not every task needs GPT-4 or Claude Sonnet.
tenant-isolation-vector-database
Multi-Tenant LLM Architecture: Decisions I Got Right and One I'd Fully Reverse
The B2B SaaS context changes almost every architectural decision in an LLM system. You are not building a product with one corpus and one user population.
langgraph-vs-crewai-vs-autogen
The Agentic AI Tools I Trust in Production (and the Ones I've Stopped Recommending)
I want to be upfront about what this post is and isn't. It's not a comprehensive market survey.
prompt-management-llm-production
How I Actually Version Prompts in Production (Not the Ideal System)
There is a version of this post where I describe a clean prompt management architecture: a dedicated prompt store, semantic versioning, A/B deployment, automatic rollback on degradation, beautiful dashboards. That system exists.
ml-pipeline-dag-checkpointing
Building the CI/CD Engine Inside Pipeshift: What the Architecture Looks Like and What I Got Wrong
I am the founder of Pipeshift, so everything in this post is written with that bias on the table. This is not an objective analysis of ML CI/CD tooling in general.
claude-code-workflow
Claude for Code: What Works and What Costs Me Time
I use Claude daily. I have built parts of Pipeshift's pipeline tooling on top of the Claude API -- I am the founder of Pipeshift, so that context matters when you weigh anything I say here.
llm-api-cost-optimization
The Real Cost of Running Frontier Models at Production Scale
The API pricing page is not your cost. That number -- $15 per million output tokens, $3 per million input tokens, whatever it is this week -- is the floor.
llm-agent-blast-radius
I Shipped an Agentic Feature Before It Was Ready. A User's Staging Environment Paid for That.
I want to be specific about what happened, because the generic version of this story -- "AI agent did unexpected things" -- gets told without enough detail to be useful.
when-to-fine-tune-vs-rag
My Decision Tree: RAG or Fine-Tune?
Every engagement starts the same way. The client has a use case -- a domain-specific assistant, a document QA system, a support bot that needs to sound like the company -- and someone on their team has already formed an opinion.
self-hosted-llm-cost-comparison
The Calculation I Run Before Sending a Request to OpenAI
I did not set out to reduce my OpenAI usage. I set out to understand what I was actually paying for, and the answer surprised me enough that three workload categories are now running on something else.
gpu-node-pool-kubernetes
How I Actually Set Up Kubernetes for ML Inference in Production
Every tutorial for deploying ML models on Kubernetes follows the same path: create a Deployment, set up a Service, maybe wire in an HPA on CPU utilization, call it done. That path is fine for getting something running in an afternoon.
llm-evaluation-pipeline
I Built an LLM Eval Harness From Scratch. Here Is What That Cost Me and What I'd Do Differently.
The honest version of "why I built my own eval harness instead of using an off-the-shelf tool" is not ideological. I did not build it because I think NIH is virtuous or because I distrust existing tools.
long-context-vs-rag
Bigger Context Windows Are Actively Harmful for Some Production Workloads
Every few months a model provider announces a larger context window as if it is a straightforward quality improvement. 200k tokens. 1M tokens. And yes, for some workloads those numbers matter.
agent-orchestration-layer-design
The Architecture Patterns I Use for Every Agentic System I Build
At 2am on a Tuesday, one of my early agent systems was stuck in a loop -- tool call, failed parse, retry, failed parse, retry -- and the retries were not bounded. It had been running for forty minutes.
mlops-platforms-worth-it
MLOps Is a Sales Category. Good ML DevOps Is Just Engineering Discipline.
The term "MLOps" started as a reasonable shorthand for "operational practices applied to machine learning systems." Somewhere between 2021 and now it became a vendor category, a conference track, a job title prefix, and a bucket of platform marketing.
rag-retrieval-quality-production
How a RAG System I Built Was Hallucinating on Every Third Query -- and What I Missed
I got a Slack message at 3:14am on a Tuesday. The client -- a B2B SaaS company, anonymized here -- had deployed a RAG-powered internal knowledge assistant I had built for them about six weeks earlier.
ragas-metrics-production
The RAG Evaluation Setup I Actually Use in 2026
The standard advice on RAG evaluation is to "use RAGAS and check your metrics." That advice is not wrong.
retrieval-vs-llm
Most "AI Features" in B2B SaaS Are Just Search With a Better Mouth
The demo looks like this: user types a question in natural language, a friendly response appears that cites specific records, the founder calls it "AI-powered." Investors nod. Engineers ship it.
qdrant-vs-weaviate-vs-pgvector
Vector Database Benchmarks I Actually Ran: Qdrant, Weaviate, and pgvector at 1M and 10M Vectors
The vector database benchmark posts I keep finding online share one characteristic: they were run by the vendors, on hardware the vendors control, against query distributions that favor their product. I don't find them useful.
building-a-developer-tool-startup
What the First 90 Days of Building Pipeshift Actually Looked Like
*Full disclosure: I am the founder of Pipeshift. Everything I write about Pipeshift is written from that position.
llm-agent-reliability
The Five Agentic AI Failure Modes I Keep Seeing in Production
Every team I review thinks they built a different system. The product names differ, the domains differ, the models differ. The failure modes are nearly identical.
jenkins-ml-pipelines
The ML Pipeline CI/CD Setup I Actually Use -- and the Failure That Forced Me to Build It Right
*Full disclosure: I'm building [Pipeshift](https://mohakdeepsingh.dev/products), a tool for managing ML pipeline deployments. The architecture I describe here is the direct predecessor to what Pipeshift automates.
rag-chunk-size-production
Section-Level vs Paragraph Chunking: I Benchmarked Both on 15k Technical Docs
The answer to "what chunk size should I use?" is always "it depends," and I've always found that answer useless. It depends on *what*, specifically, is the question. I ran a benchmark to find out.
langchain-alternatives
Why I Stopped Using LangChain
There was a specific moment. I was three hours into debugging a production retrieval failure, staring at a traceback that ran through six layers of LangChain internals before surfacing anything I could act on.
kubernetesgodevopsociawsinfrastructure
Why I Wrote a Go CLI Instead of Bash for Kubernetes Provisioning (And What Idempotency Actually Required)
The Kubernetes provisioning runbook was a bash script that had grown to several hundred lines with conditional logic for three environments across two clouds. That is the point where you rewrite it in Go. Here is what that decision cost and what it gave back.
ai-governanceenterprisesecurityrbacmcpsso
Enterprise AI Tooling Governance: RBAC, MCP Configs, and SSO for GitHub Copilot, Claude Code, and Cursor
Most enterprise AI tooling rollouts are underprepared for the governance questions. Here is the framework I use: RBAC tiers, MCP server configurations, Keycloak SSO integration, and LLM acceptance criteria for engineering workflows.
terraformociinfrastructuredevopsiac
A Week of terraform import: Recovering a Drifted OCI State File
The engagement did not start with building the OCI landing zone. It started with figuring out what actually existed in the tenancy. A week of terraform import, in dependency order, and what that methodology looked like.
ragaillmvector-searchoraclepython
Fixed-Size Chunking Breaks Structured Documents. Here Is What I Use Instead.
Fixed-size chunking is the default for RAG tutorials and the wrong choice for hierarchically structured documents. The section-level strategy I built for a production Oracle HLD generator -- and where it still falls short.
agentic-aiarchitecturesecurityaicloudllm
The Well-Architected Framework for Agentic AI: A Practical Guide
Operational excellence, security, reliability, performance, cost, sustainability — applied to AI agents that actually survive production. The rubric most teams skip.
ci-cdjenkinsgithub-actionsdevopspipeshift
How We Migrated from Jenkins to GitHub Actions (And What We Learned Building Pipeshift)
A practical account of migrating production CI/CD from Jenkins to GitHub Actions — the decisions, the tradeoffs, and the patterns that didn't survive the translation.
kubernetesociawsdevopsinfrastructure
OCI vs AWS for Kubernetes: An Honest Comparison After Running Both
After running production Kubernetes on both OKE and EKS, here's where OCI wins, where AWS wins, and where the answer is genuinely 'it depends' for real reasons.
terraformdevopsmulti-cloudociawsazure
Terraform State Strategy for Multi-Cloud Teams
Remote state, environment isolation, and the guardrails that prevent your multi-cloud IaC from becoming a liability. What I've learned running OCI, AWS, and Azure in parallel.
langgraphaiociarchitecturepythonllm
LangGraph in Production on OCI: Patterns That Hold Up
What actually breaks when you take a LangGraph agent pipeline off your laptop and run it on OCI Functions — and the patterns that survive contact with production.
kubernetescostdevopsreliabilityociaws
Kubernetes Cost Optimization Playbook
The actual levers for reducing Kubernetes spend without regressing reliability — with numbers from real clusters on OKE, EKS, and AKS.