August 9, 2026
Inside a Multi-Agent SEO Machine: 6 Months in Production
I run a production SEO workflow where specialized agents research, write, validate, and publish with strict permissions. Durable job records, versioning
Lazar Milicevic - Blog
First-person essays on AI automation, GenAI/LLM engineering, cloud architecture and building autonomous systems, by Lazar Milicevic.
August 7, 2026
Building a Sovereign AI POC: Local LLMs and RAG
I built a sovereign AI POC where no document, embedding, or model weight leaves the customer's network, using self-hosted Llama/Qwen models (32B-70B)
August 6, 2026
Hybrid Search in Postgres: pgvector + FTS with RRF
I run hybrid search on Postgres 16 with pgvector and tsvector across 1,115 posts, fusing results via Reciprocal Rank Fusion (k=60) because pure vector
August 5, 2026
Measuring AI Assistant Recommendations in Production
Across seven production sites, Gemini mentioned my brands 13 times in 32 prompts and recommended them 14 times, while Claude mentioned them 26 times in 94
August 4, 2026
Running Claude Code Agents in Production: A Field Report
I run 1,115 pieces across 7 sites and 183 local videos from one box using Postgres 16, Python workers, and Windows Task Scheduler, no queues, no
August 3, 2026
How I Stop an Autonomous AI Agent From Publishing Garbage
I run an autonomous content pipeline producing 1106+ articles across 7 sites with zero human editors, using seven pre-publish gates backed by Postgres
August 3, 2026
How to Build an AI POC That Convinces Stakeholders
I scope AI POCs around one sentence: a named decider will make a specific go/no-go call based on concrete evidence, with must-hit metrics signed off
August 2, 2026
Designing a Generative AI POC: From Prompt to Production Agents
I design generative AI POCs that survive the jump from demo to unattended production, based on what actually held up building ContentStudio. The hard
August 1, 2026
Inside an AI Implementation That Actually Lasts
I ship a serverless RAG-plus-agent pattern on AWS with hybrid retrieval, reranking, and Claude tool use, wrapped in per-tenant guardrails, schema-checked
July 31, 2026
Why Agentic AI Stalls at the Demo
Agentic AI demos fail in production because teams spend 80% of effort on prompts instead of state, retries, idempotency, and human checkpoints, turning a
July 30, 2026
From Idea to Working AI POC in 10 Days
I ship AI proofs of concept in ten days by freezing scope to one decision, one workflow, and a binary success bar before writing code. I read 100 real
July 29, 2026
AI Agent Stack in 2026: LangGraph vs Custom vs DIY
I ship AI agents to production weekly and pick from three stacks: a ~150-line DIY loop for simple linear tools, LangGraph when I need typed state
July 28, 2026
Agentic Workflows That Survive Real Inputs
I design multi-agent workflows daily and the pattern that survives production is: write the workflow as a deterministic pipeline first, then let an LLM
July 27, 2026
From Fannie Mae Systems to Autonomous AI Agents
I build autonomous AI agents with the same discipline as regulated financial batch systems: versioned prompts, run IDs, logged tool calls, typed schemas
July 26, 2026
Claude Opus 5: What Half-Price Intelligence Changes
Reported Claude Opus 5 pricing could make multi-step agents cheaper, but I would verify availability and terms with Anthropic before production use. I
July 25, 2026
What an AI MVP Actually Costs to Build in 2026
An AI MVP in 2026 costs between $12,000 for a lean proof-of-concept and $70,000+ for a production-grade system, with engineering time, not API spend
July 24, 2026
Agentic Workflows for Software Development: What Actually Works
I've spent over a year shipping production code with agentic workflows across SaaS features, serverless integrations, and autonomous content systems. Here
July 22, 2026
Building Agentic Workflows with Claude Code
I run Claude Code as the execution layer for autonomous systems like BizFlowAI ContentStudio, and reliability came from workflow design, not smarter
July 21, 2026
Cutting LLM Token Spend 40% Without Losing Accuracy
I cut production LLM costs 40% by routing 60-70% of requests to cheaper tiers (local, Haiku, or mini models) and reserving flagship models for real
July 18, 2026
Millisecond Agents, Second-Scale APIs: Closing the Gap
At VB Transform 2026, infra leaders confirmed what I see daily: models aren't the bottleneck, the 2-second legacy APIs behind them are. I close that gap
July 17, 2026
The Agent Security Gap Nobody Wants to Talk About
A survey of 107 enterprises found 54% had an agent security incident or near-miss, yet most still share credentials across agents, collapsing attribution
July 16, 2026
Why 95% of AI Agent Pilots Never Ship
Most AI agent pilots stall because per-step model reliability compounds into system-wide failure, not because models aren't smart enough. I ship agents to
July 15, 2026
One API Key, Five Agents: How to Stop the Blast Radius
Shared API keys across agents collapse the permission model and destroy forensic traceability, turning any prompt injection into a union-of-all-scopes
July 14, 2026
Building an AI SaaS MVP in 2026: My Stack and Timeline
I ship AI SaaS MVPs on a 4-week clock using Next.js 15, Supabase with pgvector, Claude Sonnet with GPT-4.1 fallback, and Stripe usage-based billing. The
July 13, 2026
RAG Eval Metrics 2026: Ragas vs TruLens vs DeepEval
I've spent three years running RAG systems in production, and this guide compares the three eval frameworks I actually use, Ragas, TruLens, and DeepEval
July 12, 2026
The Agentic Context Layer: Fixing Confidently Wrong Agents
I've hit production RAG failures where agents confidently served stale numbers because definitions had drifted across docs and retrieval never caught it
July 11, 2026
The Enterprise AI Evaluation Gap Is Widening
Half of enterprises have shipped AI features that passed internal evals but still failed in production, because most eval suites test the model in
July 10, 2026
Projects Built With Claude Code: Real Systems in 2026
In 2026 I shipped three production systems with Claude Code driving most of the code: a multi-agent SEO pipeline, serverless AWS support glue, and a
July 9, 2026
What Billions of Predictions Taught Expedia About AI at Scale
Running AI in production isn't about shipping the model, it's about surviving month six when data, prompts, and model versions have all silently drifted
July 7, 2026
How to Build AI Agents From Scratch: A Blueprint
I strip frameworks like LangChain and LangGraph out of production because an agent is really just a while-loop around an LLM call with tools, a hard step
July 6, 2026
LLM Evals in 2026: A Practitioner's Field Guide
LLM evals in 2026 span offline benchmarks, online traces, guardrails, and judge calibration, and most teams conflate them. I share the 12 failure modes I
July 5, 2026
n8n in Production: From First Node to Self-Running AI Workflows
I share what I've learned running n8n in production for real AI workloads, including the architecture I use for autonomous content pipelines and the error
July 5, 2026
RAG in Production: A Senior Engineer's Architecture Guide
I share hard-won lessons from building RAG pipelines in production: why chunking strategy matters more than embedding model choice, how reranking delivers
July 5, 2026
What Is an AI Agent? A Working Engineer's Definition for 2026
An AI agent is an LLM with tools, memory, and a control loop that lets it decide its own next steps until a task is complete. I define what separates
July 5, 2026
7 Characteristics of Production-Grade AI Agents (And Why Most
I run multi-agent pipelines in production, and these seven characteristics determine whether an AI agent survives unattended or only works in demos. They
July 4, 2026
AI Proof-of-Concept Examples That Actually Made It to Production
I share the AI proof-of-concepts that actually made it to production for me, including a multi-agent content automation system, a hybrid RAG pipeline with
July 4, 2026
What Real Generative AI Consulting Looks Like in 2026
I run GenAI consulting engagements by killing bad projects in scoping, then making structural architecture decisions, sync vs. batch, single-call vs
July 4, 2026
LangChain in Production: What I Keep and What I Rip Out
I keep LangChain's document loaders, text splitters, prompt templates, and output parsers because they handle the boring plumbing reliably. I rip out the
July 3, 2026
118K vs 3.26M Tokens: Memory Design for Agents
Long-horizon agents like MRAgent hit 118K tokens per query versus 3.26M for LangMem because most teams treat context as append-only storage instead of
July 3, 2026
Claude Code turned every engineer into three
Claude Code roughly tripled my shipping speed on backend and integration work, but the bottleneck moved from writing code to specifying work, reviewing
July 3, 2026
Cutting Agent Token Use 99% With Tool Routing
I cut a 147-tool agent's input from 38,000 tokens to a fraction by routing tools through pgvector retrieval instead of registering all schemas up front. A
July 2, 2026
LLM Evals: Hamel's Method vs a Lean In-House Setup
I've run both LLM eval paths on production systems, and the honest tradeoff is this: a lean in-house setup ships in a week for ~$200 and catches 70% of
July 2, 2026
50 AI Implementations That Save Time and Money
I score 50 AI implementations across sales, support, and operations on effort and value (1-5), drawing on projects I've built or deployed, including a
July 1, 2026
Hiring an AI Automation Consultant for Your Small Business
I help small businesses under 50 people replace their most painful manual workflows, document extraction, CRM updates, ticket routing, with working
July 1, 2026
What Is an AI Automation Consultant? A Real Definition
BizFlowAI's founder defines an AI automation consultant as someone who designs, builds, and hands over autonomous systems that replace repetitive work
July 1, 2026
DSpark vs vLLM: What an 85% Speedup Really Means
I benchmarked DeepSeek's new DSpark against vLLM on Llama-3.1-70B and found the "85% faster" claim only holds at large batches and long generations, real
July 1, 2026
Less Autonomy, More Value: The Morgan Stanley Lesson
Morgan Stanley cut daily P&L reconciliation work in half by making their AI agents less autonomous, not more, agents investigate and package cases, but
June 30, 2026
15 Real AI Implementations and What They Shipped
I've shipped 15 production AI systems spanning RAG, agentic workflows, and serverless automation, and the pattern that pays is boring: structured
June 30, 2026
How to Run an AI POC That Ships to Production
I rescue AI POCs that stall because no one defined "done" before coding, demos shine on cherry-picked inputs, then production is six months away. My fix
June 30, 2026
Agentjacking via Sentry: How I Sandbox Claude Code
I sandbox Claude Code after reproducing the Tenet Sentry agentjacking attack, where a single crafted error event posted to a public DSN got the agent to
June 29, 2026
Prompt Injection: The Threat I Plan For First
I design every enterprise AI system assuming prompt injection will succeed, because LLMs can't reliably separate instructions from data. My defense is
June 28, 2026
The AI Implementation Process I Use With Every Client
I've refined a five-phase AI implementation process, scoping, POC, integration, evaluation, operations, where each phase has a hard exit criterion that
June 27, 2026
How to Create an AI Agent: A Production Walkthrough
After a 3am production failure cost me $40 in a tool-call loop, I rebuilt my agent pattern around tight job specs, well-designed idempotent tools, and
June 26, 2026
Token Rationing: Guardrails for AI Spend
Token spend explodes when teams treat AI like an unlimited resource — I've seen six-figure monthly bills where 40% was devs asking trivial questions. The
June 26, 2026
ML Platform Build vs Buy in 2026: Real TCO
I built a real TCO model across Databricks, SageMaker, Vertex AI, and self-hosted Ray+MLflow at three company sizes, with 2026 list pricing. The honest
June 26, 2026
llm CLI vs Aider vs Cursor CLI vs Claude Code: 2026
I run all four tools across different machines—Claude Code for autonomous multi-file work, Aider as a git-disciplined pair-programmer, Simon Willison's
June 26, 2026
Building AI Agents That Safely Query Enterprise Databases
I built guardrails for AI agents querying production databases after one tried to SELECT * a 400M-row table—starting with a read-only Postgres role
June 26, 2026
AI System Cost in 2026: Real Budget Breakdown
I budget AI builds in 2026 across three distinct stages: PoCs at $8k–$25k, MVPs at $40k–$150k, and production systems at $150k–$600k+. Roughly 70% of cost
June 26, 2026
The Fractional AI Forward Deployed Engineer
I work as a fractional forward deployed AI engineer, embedding 1-2 days a week with multiple SMBs to ship RAG pipelines, agents, and evals directly in
June 25, 2026
PoC to Production: Scale AI Without a Rewrite
PoCs usually fail not because of model quality but because day-one architectural shortcuts become load-bearing by month six. I lock in four things early —
June 25, 2026
RAG Evaluation Metrics: A 12-Point Checklist (2026)
I rebuilt my RAG evaluation after a pipeline scored 0.91 on answer relevancy while hallucinating account numbers in production, and now run a 12-metric
June 25, 2026
LLM-as-a-Judge: Build, Calibrate, Trust It
I shipped an LLM judge that scored outputs 4.6/5 when humans rated them 3, so I built a calibration recipe using a 150-sample human-labeled gold set
June 25, 2026
Lazar Milicevic vs Hamel Husain: LLM Eval Approaches
Most engineers ask how my LLM eval approach differs from Hamel Husain's: we agree on fundamentals like error analysis and sparing LLM-as-judge use, but
June 24, 2026
How I Build Systems That Run While I Sleep
I build unattended systems on four properties in order—scheduling, idempotency, observability, and graceful failure—starting with boring scaffolding
June 24, 2026
What an AI Automation Engineer Actually Does
I build unattended systems that replace recurring manual work—mapping the real process, automating deterministic parts with code and judgment calls with