Writing
Technical deep-dives on AI architecture, RAG systems, voice AI, enterprise software, and 14+ years of building.
Deep dive into production AI agent architectures — ReAct, Plan-Execute, Multi-Agent — with real examples from Hureka AI and AImind. Covers tool calling, memory, monitoring with LangFuse, and battle-tested patterns.
Complete guide to building production RAG systems — chunking strategies, embedding models, hybrid search with Qdrant, reranking, evaluation metrics, and deployment patterns for enterprise scale.
LangGraph adds graph-based state machines to LangChain. Learn how to model multi-agent coordination, conditional branching, human-in-the-loop, and persistent state for production AI workflows.
Detailed cost comparison and architecture guide for self-hosted Voice AI using Pipecat, LiveKit, and Whisper vs cloud solutions like Twilio AI. Real production metrics and latency optimization.
Practical guide to building AI SaaS products — multi-tenancy patterns, BYOK (bring your own key), billing integration, feature flags, and deployment strategies with lessons from building Hureka AI.
A practical comparison of the four most-used vector databases for production RAG — covering latency, cost, hybrid search, filtering, self-hosting, and which one I pick in each scenario.
Comprehensive LLMOps guide covering model serving with Ollama and vLLM, caching strategies, prompt versioning, monitoring with LangFuse, cost tracking, A/B testing prompts, and rollback patterns.
Honest comparison of hiring AI architects through Upwork, Toptal, and direct engagement. Covers hidden costs, quality differences, real project outcomes, and how to evaluate genuine AI talent.
Enterprise-focused comparison of Anthropic Claude vs GPT-4 — context windows, safety features, pricing, and real use cases. Includes multi-LLM strategies for production applications.
Production prompt engineering is software engineering. Learn how to template, version, evaluate, and A/B test prompts so your AI features improve continuously instead of regressing silently.
Battle-tested strategies to reduce AI infrastructure costs — self-hosting vs cloud comparison, semantic caching, model distillation, batching, prompt optimization, with real production numbers.
Comprehensive guide to AI consulting engagements — what an architecture review covers, engagement models, timelines, deliverables, pricing transparency, and red flags to watch for.
A complete guide to horizontally scaling WebSocket servers using Redis Streams as the pub/sub backbone. Connection draining, sticky sessions, message ordering, and observability.
How I reduced LLM costs for production AI products from $42K/month to $12K/month without sacrificing quality. Caching, routing, distillation, prompt compression, and more.
How to design LLM tool/function schemas that actually work in production. Strict schemas, parallel calls, error recovery, multi-step reasoning, and the failure modes nobody warns you about.
A complete guide to running Ollama in production. GPU selection, concurrent request handling, model warmup, quantization choices, and the gotchas that take down hobby setups when real traffic hits.
The three multi-tenant database strategies for SaaS — shared schema with tenant_id, schema-per-tenant, and database-per-tenant. When to use which, with real migration paths between them.
How to architect memory for AI agents that need to learn from past interactions. Short-term context windows, long-term vector memory, episodic memory, and semantic distillation patterns.
VAD decides when the user is done speaking. Get it wrong and the agent interrupts or hangs. A deep dive into Silero VAD, energy thresholds, end-of-turn detection, and barge-in handling.
How to evaluate RAG quality rigorously. Faithfulness, answer relevance, context precision, context recall — using Ragas to catch regressions before users do.
How to build a production chat interface with React Server Components in Next.js 15 — streaming LLM responses, optimistic UI, markdown rendering, code highlighting, and message persistence.
Battle-tested Go patterns for microservices — context propagation, graceful shutdown, structured logging, circuit breakers, gRPC + HTTP duality, and the boring infrastructure that wins.
The right way to do JWT authentication in Next.js 15 with App Router — middleware, refresh tokens, secure cookies, role-based access, and server components that know who you are.
Building AI that assists clinicians without overstepping. Decision support, not diagnosis. Architecture patterns, guardrails, audit trails, and how to design for the liability questions that always come.
What I learned managing a fully-remote AI engineering team across 5 time zones. Hiring, async culture, on-call rotations, technical reviews, and the soft skills that matter more than the framework wars.
How to instrument production LLM applications with LangFuse. Traces, scoring, cost attribution, prompt management, and the dashboards that let you ship AI confidently.
When pgvector beats a dedicated vector database. Index choices (HNSW vs IVFFlat), tuning for 10M+ rows, hybrid search with full-text, and the moment you should reach for Qdrant instead.
Generic embeddings (BGE, OpenAI) leave 20% recall on the table for domain text. Learn how to mine training pairs from your own documents and fine-tune sentence-transformers for medical, legal, or financial RAG.
How I migrated a 9-year-old Drupal 7 site with 80K nodes, 200K users, and 14 years of SEO into a modern Next.js + headless CMS architecture without losing rankings or breaking URLs.
A practical guide to squeezing maximum performance from Next.js 15 — React Server Components, granular caching strategies, streaming Suspense boundaries, image optimization, and hitting 100 on Lighthouse.
How to design AI-powered healthcare systems that pass HIPAA, SOC 2, and GDPR audits. Real patterns from building DrMackMedicine — covering PHI handling, audit logging, and compliant LLM usage.
The architecture patterns that take FastAPI from a quick prototype to a production-grade enterprise API — dependency injection, background tasks, streaming responses, multi-tenant middleware, and deployment.
A practical guide to designing multi-agent AI platforms with shared RAG brains, Qdrant vector databases, and FastAPI backends. Lessons from building AImind and Clinic AI at Hureka Technologies.
How to create production-ready Docker images under 150MB for Next.js (standalone output) and FastAPI. Multi-stage builds, layer caching strategies, non-root users, and health checks.
How to build a fully self-hosted voice AI agent with zero cloud dependencies. Pipecat orchestrates Faster-Whisper STT, Ollama LLM, and pyttsx3 TTS over LiveKit WebRTC.
Everything I learned building production RAG systems. Optimal chunk sizes, embedding model selection, Qdrant HNSW tuning, hybrid search, and reranking strategies.
Temporal gives AI applications durable execution, automatic retries, and observable state. Learn how I use Temporal for multi-step LLM workflows, email automation, and BYOK AI SaaS.
How I architected Hureka AI — a BYOK (Bring Your Own Key) multi-tenant AI SaaS. AES-256 encryption for API keys, and a unified LLM adapter pattern for Anthropic, OpenAI, Google, and Ollama.
My journey from junior web developer to Lead AI Architect — how enterprise CMS work shaped my thinking about AI systems, and the principles that never change regardless of technology.
39 posts · More coming soon
