AI Transformation Lead · Agentic AI · Forward Deployed Engineering
Pooya Golchian is an AI Transformation Lead in Dubai. He builds the harness that decides whether an agent's work is safe to ship.
I work inside your repo, not beside it. Coding agents write most of the code now, and your engineers own the harness the day I leave.
Shipping software since 2009, the last three years of it in production AI. Read the background and founder roles on About Pooya Golchian, or start smaller and audit your own stack against the free 14-point agent readiness checklist.
Agentic software engineering is not a buzzword here. It is a job I do inside your team, a documented lifecycle, a way of working with clients, and a real set of Claude Code skills and agents I run every day.
What the role is, what it demands day to day, and the path into it.
MethodologyThe AI Development Life Cycle. Eight phases from a framed problem to an eval-guarded system.
ServicesAgentic marketing, development, and systems integration. Production work, not pilots.
ToolkitThe skills, agents, and plugins I build to do the work. Filter by type or browse the full set.
Senior engineering for production AI, private LLMs, full-stack systems, and the teams that operate them.
Production agent systems, MCP servers, and RAG. Senior engineering for teams shipping AI into real apps.
Tool access, specs, evals, and guardrails around your agents. Deploy on infrastructure you control when residency rules require it.
Production Next.js, Node.js, and monorepo work. Server Components, durable workflows, performance budgets.
Hire, train, and structure a senior AI engineering team. Recruitment, interviews, onboarding, and operating cadence.
Train an existing engineering team to ship with agents. Workflow assessment, hands-on AIDLC training, and a rollout plan managers can run.
Migrate from Vercel or Firebase to AWS. Cut hosting costs 40-70% with SST v3, Lambda, CloudFront, and DynamoDB.
Growth and lifecycle work stays available when it supports an AI product or B2B operator workflow.
Members-only writing on how agentic software actually gets built. AIDLC, bolts and specs, AI team building, AWS BFF and serverless, agentic programming, and the skills and plugins behind the workflow.
Members only. Sign in to read the full library.
Working notes from shipping agentic AI into production. Agent systems, MCP servers, RAG that cites its sources, evals that catch drift before your customers do, and the AIDLC method I use to move B2B teams past the pilot stage.
One Google Maps query stops at 200 results, so enumerating a city takes a different shape. Here is what 1,400 SearchApi requests actually bought me.
Read
SpecForge fuses spec-kit's executable specs with AI-DLC governance, then enforces approval with a Claude Code hook that blocks code edits until you sign off.
Read
When agents write most of the code, review stops being a checkpoint and becomes the job. A members-only playbook for building a verification loop that keeps trust high without slowing the team to a crawl.
Members only, open in the Lab
A bolt is a short spec, generate, verify cycle that a team runs in hours instead of a week-long sprint. A members-only playbook for intake, roles, cadence, and the metrics that keep it honest.
Members only, open in the Lab
The two-week sprint was built around human typing speed. When agents write the code, the bolt replaces it. A side-by-side on planning, cadence, roles, and the metrics that change.
Read
A practical guide to building an app by pointing a harness at a goal prompt and running it in a loop: how to write the goal, wire the verification gate, and know when to stop.
Read
A spec is the highest-leverage ten minutes in an agentic cycle. A practical guide to writing specs that stop intent drift, from acceptance checks to sizing for a single bolt.
Read
Buying AI seats does not make a team agentic. In 2026 the bottleneck moved from writing code to verifying it. Here is how to retrain the workflow, the people, and the way you manage them.
Read
Specs and bolts are two halves of one cycle. How the spec-to-bolt loop runs end to end, from intake to verified merge, and why it scales to many agents at once.
Read
A bolt is the AIDLC unit of work that replaces the two-week sprint with a cycle measured in hours. What a bolt is, why AWS renamed the sprint, and how teams run one.
Read
An agent harness is the deterministic layer that wraps an LLM: the loop, the tools, the sandbox, the permissions. Why the harness, not the model, is the 2026 differentiator.
Read
Spec-driven development treats a written specification as the executable source of truth and code as a regenerable output. What SDD is, why it beat vibe coding, and the tools that ship it.
Read
The Ralph loop runs a coding agent in a plain while loop against one goal prompt, with memory in files and git instead of context. What it is, why it works, and where it breaks.
Read
A 2026 AI transformation playbook for UAE and Dubai firms covering a phased roadmap, real ROI, PDPL and DIFC compliance, and where a consultant fits.
Read
AIDLC vs SDLC compared phase by phase. Where the effort shifts from implementation to verification, which phases shrink, and what that costs a team mid-adoption.
Read
Ollama Cloud starts at $0 and caps at $100/mo flat. Claude Sonnet runs $3/$15 per 1M tokens. Where each wins on cost, limits, and quality in 2026.
Read
Run local LLMs with GitHub Copilot in Visual Studio 2026 via Ollama. BYOK setup, model choices, endpoint wiring, and where local beats cloud.
Read
Agents can write the ETL. What they cannot do is decide what an agent should retrieve, how to ground it in your data, and which sources are trustworthy enough to feed a decision. Data engineering moved to retrieval and ground truth.
Read
When part of your team is human and part is autonomous, velocity metrics lie and the old planning rhythm breaks. The engineering manager's job moved to designing the operating model where people and agents ship together.
Read
Agents draft docs in seconds, so writing prose is no longer the scarce skill. The technical writer's value moved to authoring the specs and context that agents read, and to making sure both humans and machines can act on it.
Read
Prompt injection, tool abuse, and data exfiltration through an agent are not edge cases, they are the new baseline. The security engineer's job expanded from securing code to securing autonomous decision-making.
Read
Agents can write the Terraform and the CI config. What they cannot do is decide what an autonomous system must expose to stay safe in production. DevOps moved from automating pipelines to instrumenting agency.
Read
Deterministic test suites cannot judge probabilistic systems. The QA engineer's role moved from asserting exact outputs to designing evals, golden datasets, and LLM-as-judge suites that gate agentic behaviour.
Read
Agents scaffold components, wire state, and match a design in minutes. The frontend developer's edge moved to design systems, accessibility, performance budgets, and the taste to reject output that merely looks done.
Read
Agents now write most of the CRUD, the handlers, and the glue. The backend developer's value moved to data contracts, correctness review, and the judgment to catch the bug the agent shipped with confidence.
Read
Agents generate code fast, including the wrong abstraction fast. The architect's job shifted from drawing diagrams to designing the boundaries, contracts, and scaffolds that keep agent output inside the lines.
Read
When agents write the code, the product manager's leverage moves upstream. The PMs who win in 2026 author executable specs, own the success metric, and frame problems agents can act on. Here is how the role changed.
Read
Google's Gemma 4 update ships Multi-Token Prediction for roughly 3x faster decoding. Benchmarks against Qwen 2.5 and Llama 3.3, exact ollama pull commands, and the hardware floor.
Read
Engineers complain that AI hallucinates, burns tokens, and ships nothing. Each of those failures has a named technique that fixes it. The frustration is a skills gap, and the fix is a working method.
Read
Most B2B companies hire for AI and end up with a single ChatGPT wrapper. The agent teams that ship use a specific architecture for roles, memory, tools, evals, and deployment. Here is the process I run with clients in 2026.
Read
riyal v1.2.0 ships the Saudi Riyal symbol (U+20C1) as a web font, React, Vue 3, Svelte 5, React Native, and Web Component primitives. Includes a masked SAR input, cart math with 15% VAT, a Tailwind plugin, OG cards, a CLI, and a shadcn registry. Built off the playbook that took the dirham package to 8K monthly downloads.
Read
Engineers blow through Claude Max quotas in four days because tool output floods the context. RTK is a Rust proxy that filters command output at source and reclaims 60-90% of every coding session.
Read
dirham v1.5.3 ships animated price counters, masked currency inputs, live exchange rate hooks, a Tailwind plugin, a Next.js font helper, and React Native support. Full implementation guide with code examples.
Read
Claude Projects, GPT memory, and Gemini context windows compared. How each AI maintains state across conversations, token limits, and implementation patterns for production agents.
Read
Teams using AI code review tools ship 40% faster with equivalent bug rates. Implementation patterns, tool comparison, and metrics from companies like Shopify, Vercel, and Linear.
Read
Anthropic's Claude Max subscription costs $350/month but enables 20x more usage than Claude Pro. Benchmarks from senior engineers who switched: code review velocity, architectural decision speed, and documentation generation at professional scale.
Read
I fired 500 requests per tier at five headless CMS from us-east-1 in April 2026. The DX chart and the price table name different winners.
Read
CrewAI pricing for 2026 next to LangGraph, AutoGen and Smolagents, plus 200 benchmark tasks each on Qwen3 32B. The cheapest plan rarely wins on cost.
Read
Hermes Agent from Nous Research runs entirely local through Ollama with 70+ built-in skills, cross-session memory, and messaging gateway integrations. Setup guide, model benchmarks, and comparison with Claude Code for autonomous workflows.
Read
Ollama revised Cloud quotas twice since GA. Three tiers measured against a 36-month amortized RTX 4090 and Mac Studio, plus the volume where each flips.
Read
Benchmarks comparing Rust, Go, and Zig for backend services. Memory safety, concurrency models, compilation times, and production experiences from teams that migrated.
Read
TypeScript 6.0 npm downloads jumped 12.5% in two weeks after release. The trending story is not about new syntax. It is about erasable types, native NodeJS support dropping, and a decade of promises finally shipping. Here is what changed.
Read
Claude Opus 4.6 leads on agentic coding with computer use capabilities. GPT-5.3-Codex excels on terminal operations and web development. A detailed benchmark comparison for engineering teams choosing AI coding tools.
Read
Google Gemini 2.0, OpenAI GPT-5.3, and Anthropic Claude Opus 4.6 represent the current frontier. A comprehensive benchmark comparison across coding, reasoning, multimodal, and agentic capabilities for engineering team AI tool selection.
Read
Ollama's top HumanEval score runs at 8 tokens per second. Four VRAM tiers, an RTX 5070 baseline, and break-even math at 500 queries a day.
Read
Claude Code configuration options that separate junior from senior usage. Custom instructions, workspace modes, tool restrictions, and workflow patterns that unlock maximum productivity for experienced developers.
Read
Explore the fundamental principles of design including balance, contrast, emphasis, movement, pattern, rhythm, proportion, unity, and white space through cinematic AI-generated black and white photographs.
Read
Chain-of-thought reasoning models demonstrate emergent reasoning capabilities. Understanding how reasoning emerges, when it fails, and what it means for AI problem-solving in production systems.
Read
A $500 RTX 5070 running Qwen 3.5 Coder 32B outperforms Claude Sonnet 4.6 on HumanEval at 40 tokens per second. The local AI revolution has reached consumer hardware.
Read
Most teams commit it by accident. Every path Claude Code writes, the 30-day cleanup command, and how its memory holds up against Copilot and Cursor.
Read
GitHub's Octoverse 2025 data plus interviews with 22 advanced AI users, charted. The language rankings flipped. So did the job description.
Read
GitHub Copilot now runs agentic workflows through Ollama. Deploy Qwen, DeepSeek, and Llama models locally. Zero latency, complete privacy, no API costs. Full setup guide with benchmarks.
Read
7,300 open PM roles. 67,000 engineering openings. A 340% surge in AI positions. The data on who is actually getting hired in early 2026.
Read
A real-world case study of using AI to rewrite JSONata in a single day, reducing infrastructure costs by $500K per year. Learn the methodology, tools, and lessons learned.
Read
AI-Scientist-v2 timed against human researchers across six research phases, four agents, four domains. Where the autonomy claim stops holding.
ReadUAE Dirham symbol U+20C3 is now official. dirham v1.3.0 renders it today in React, Web Components, and vanilla JS with zero migration when OS support ships.
Read
GitHub rewrote Copilot's data terms in 2026. Which tiers it touches, the opt-out click path GitHub buries, and five assistants compared.
Read
Federal Reserve posting data charted from Q1 2023 to Q1 2026, seven sectors and six skills ranked. The headline number hides which developers get hired.
Read
Recent supply chain attacks targeting LiteLLM and Trivy expose critical vulnerabilities in developer tooling. Learn how these attacks work and how to protect your projects.
Read
Three charts trace build time, memory, and inference accuracy from TS 5.0 to 6.0, plus a five-step upgrade path and the tsconfig options that break.
Read
NVIDIA Nemotron 70B achieves 88.4% MMLU at 5.4x lower cost than GPT-4o. OpenManus reaches 72K GitHub stars, capturing 89% of Manus performance at 12% the cost. How open source AI agents are reshaping the $31.4B market and driving semiconductor stock valuations.
Read
Local stdio MCP servers break once SSE traffic meets a load balancer. Sticky sessions, circuit breaker thresholds, and a charted three-year cost model.
Read
From pattern matching to measured reasoning. Benchmarks show large gains from chain of thought prompting and frontier model accuracy.
Read
Cloud inference bills climb forever. Hardware bills stop. Five charts plot 18 months of TCO, benchmark four GPUs, and mark the break-even month.
Read
Four agent frameworks, five local models, three tool-use benchmarks run on Ollama. GitHub stars rank them one way. The benchmark data ranks them another.
Read
How I built a full Revenue Operations platform using Notion as the database, Gemini 2.5 Flash as the AI agent, and the Model Context Protocol to wire 22 Notion tools together without a single hardcoded query.
Read
72% of enterprises run RAG in production. Qdrant hits 6ms p50 latency. Hybrid search boosts recall 17%. Compare Pinecone, Qdrant, Weaviate, ChromaDB with embedding model scores, chunking benchmarks, and deployment patterns.
Read
$9.2B Cursor valuation. 92.4% HumanEval score for Claude 3.5 Sonnet. 340% enterprise adoption growth. Compare pricing, benchmarks, and productivity data for every major vibe coding platform in 2026.
Read
Zero per token is not zero cost. I benchmarked seven open-weight models on MMLU, HumanEval, and MT-Bench in March 2026, then ran the amortization math.
Read
Deep dive into uncertainty-aware AI: Monte Carlo methods, Bayesian inference, conformal prediction, and the statistical foundations shaping the future of trustworthy machine learning.
Read
Build AI-powered React apps that talk directly to your local models—no backend required. Stream from Ollama, LM Studio, or llama.cpp with zero dependencies and 2.8 KB of code.
Read
Headline AI risk numbers measure capability, not deployment. All 22 occupation categories from the Anthropic Economic Index, plus the hiring regression.
ReadThe dirham npm package brings the UAE Dirham currency sign (U+20C3, Unicode 18.0) to production via a custom web font, React SVG component, CSS utility class, and JavaScript formatting utilities. Drop the font import when native OS support ships in September 2026.
Read
scss-helper v5 is a modern SCSS/CSS utility toolkit that complements Tailwind CSS v3/v4 with design tokens, fluid clamp() typography, container queries, dark mode, golden ratio layouts, a 12-column grid, and animations. All classes use the sh- prefix. Only 5.1 KB gzipped.
Read
A deep dive into vue-multiple-themes v4, the open source library for dynamic CSS custom property theming in Vue 2.7+ and Vue 3. Covers the useTheme() composable, TailwindCSS plugin, WCAG contrast utilities, and dynamic theme generation from a single brand color.
Read
Introducing vue-js-star-rating, a WCAG 2.2 accessible, TypeScript-first star rating component for Vue 3.5+. Half-star precision, multiple icon providers, full keyboard navigation, read-only mode, RTL support, and zero runtime dependencies.
Read
UAE Decree-Law 45 and DIFC Regulation 10 penalize public AI usage. This guide shows how on-premise AI keeps your Dubai business compliant and audit-ready.
Read
Headless is commoditized. The real split runs developer frameworks against enterprise orchestrators. Six platforms compared on architecture, AI, and TCO.
Read
WEF Future of Jobs 2025 data: 78M new roles, 92M displaced, 39% skills gap. See which jobs grow, which vanish, and what skills matter most through 2030.
Read
Fix unnecessary React re-renders: master useCallback, useMemo, React.memo, and the new React 19 use() hook with real-world patterns, profiling tips, and pitfalls to avoid.
Read
Next.js 15, Astro 5, and Remix compared on performance, bundle size, and real use cases. Includes a decision framework to pick the right one for your project.
Read
Complete AWS ECR guide: authenticate with OIDC, pull Docker images, extract filesystem layers, scan with Amazon Inspector v2, set lifecycle policies, and automate with GitHub Actions.
Read
Production-ready shell scripts for ArangoDB on AWS EC2: automated install, daily S3 backups via systemd timers, and one-command point-in-time restore.
Read
Turn any SSH server into a private VPN with one command. Includes toggle scripts for macOS/Linux, Bitvise setup for Windows, and sshuttle vs WireGuard comparison.
Read
Step-by-step Firebase Cloud Messaging v10 setup with Vue 3: VAPID keys, service workers, foreground/background notifications, and Node.js server sending.
Read
Deploy Jekyll with a GitHub Actions CI/CD pipeline, configure a custom domain with HTTPS, and add Cloudflare CDN for speed and DDoS protection. Complete 2026 guide.
ReadThe npm registry carries 6 of them. The rest install as a Claude Code plugin or clone straight from GitHub, and every one is documented and licensed MIT or ISC.
Spec-driven AI development lifecycle with hook-enforced approval gates
Open-source spec-driven AI development lifecycle that fuses Spec Kit and AWS AI-DLC with hook-enforced approval gates.
TypeScript pipeline that turns a Google Maps crawl into a local directory
A TypeScript pipeline that turns a tiled Google Maps crawl into a browsable local directory, with plan, crawl, classify and load as separate CLIs. The Dubai reference crawl issued 1,400 SearchApi requests and returned 15,246 unique businesses with zero errors, 14,981 of them inside the city. Porting the crawl to another city is a JSON file in data/cities, with no code change anywhere in the pipeline.
AI Sales CRM built on Notion with Gemini MCP
Revenue Operations platform where Notion is the database and Gemini 2.5 Flash is the AI sales agent. Kanban pipeline, AI lead scoring, pre-call briefings, and natural language deal management — all powered by 22 Notion operations through the Model Context Protocol.
React hooks for streaming local LLM responses
React hooks for streaming responses from local LLMs — Ollama, LM Studio, llama.cpp, and any OpenAI-compatible endpoint. Direct browser-to-localhost streaming with no server, no API routes, and zero runtime dependencies beyond React.
$ pnpm add use-local-llmDynamic multi-theme support for Vue 2 & 3
Dynamic multi-theme support for Vue 2.7+ and Vue 3 via CSS custom properties. Ships with 7 preset themes, a TailwindCSS plugin, WCAG contrast utilities, a reactive composable API, and zero runtime dependencies beyond Vue.
$ pnpm add vue-multiple-themesZero-dependency Vue 3.5+ star rating component
WCAG 2.2 accessible star rating for Vue 3.5+ — TypeScript-first, half-star precision, Lucide/FontAwesome/custom icon providers, full keyboard navigation, and zero runtime dependencies.
$ pnpm add vue-js-star-ratingUAE Dirham (U+20C3) — React, Web Components, font, utilities
The UAE Dirham currency symbol (U+20C3, Unicode 18.0) as React components, Web Components, a web font, CSS utilities, and formatting helpers. Ships animated price counters, masked currency inputs, live exchange rate hooks, a Tailwind plugin, a Next.js font helper, React Native components, OG card generation, and a CLI. 1,278 weekly downloads.
$ pnpm add dirhamSaudi Riyal (U+20C1) — React, Vue, Svelte, Web Components, font, cart math
The Saudi Riyal currency symbol (U+20C1) as a web font, React, Vue 3, Svelte 5, React Native, and Web Component primitives. Ships a masked SAR input, cart and checkout math with 15% Saudi VAT defaults, currency conversion, a Tailwind v3/v4 plugin, a Next.js font helper, OG cards, a CLI, and a shadcn-compatible registry. Built off the same playbook that took the dirham package to ~8K monthly downloads.
$ pnpm add riyalModern SCSS/CSS utility toolkit filling Tailwind's gaps
A modern SCSS/CSS utility toolkit complementing Tailwind CSS v3/v4 — design tokens, fluid clamp() typography, container queries, dark mode, golden ratio layouts, animations, and a 12-column grid. All classes use an sh- prefix. Only 5.1 KB gzipped.
$ pnpm add scss-helperThe open source page adds twelve months of npm download charts, running star and download totals, and the full SpecForge breakdown.
Open source, with live download countsI design and ship production AI agent teams for B2B companies. Discovery, agent role design, MCP integration, evals, token economics, and deployment.
I join your engineering team and build the agent layer alongside you, covering architecture, MCP integration, evals, and production deployment. When the engagement ends, your team owns the system and keeps shipping.
Senior engineer turned AI specialist. React, Next.js, AWS, agent orchestration.
Direct collaboration across UAE, Europe, and US time zones.
Discovery, role design, MCP integration, evals, and production deployment.
Common questions about Pooya's work, AI services, and how to start a project together.
Pooya Golchian is an AI Transformation Lead and AI Product Engineer based in Dubai, UAE who takes B2B teams from AI strategy to production. He has shipped software since 2009, with the last three years in production AI, and his work spans agent harness design, MCP servers, coding-agent workflows, RAG and context engineering, LLM eval suites, self-hosted open-weight models, and agentic AI shipped inside Next.js and Node.js stacks. He founded Hisabi.ai and Technova Solutions, and he publishes SpecForge, an open-source method for building software the agentic way.
Pooya Golchian offers AI Engineering (production agent systems, MCP servers, RAG, eval frameworks), Agent Harness & Deployment (scoped tool access, eval suites, guardrails, and deployment on infrastructure you control with PDPL and DIFC compliance), AI Team Building (recruitment and structure), and Full-Stack Next.js and Node.js engineering with monorepo experience. He works with Claude, GPT, Gemini, Llama, and Mistral. Available to businesses in Dubai and remotely worldwide.
Yes. Pooya Golchian is available for freelance, consulting, and remote engagements. Based in Dubai, UAE, he works with clients across the Middle East, Europe, and Asia. He is open to project-based work, monthly retainers, and technical co-founding. Contact him at pooyagolchian.com/contact/ or hello@pooyagolchian.com. He typically replies within 24 hours.
Most agencies hand over a demo agent and a slide deck. Pooya Golchian builds the harness around the agent, the part that decides whether it survives production. That means scoped tool access through MCP servers, specs an agent works from instead of a prompt it improvises against, an eval suite with a golden dataset that runs on every change, cost and rate guardrails per role, and an audit log your security team can actually read. The models themselves are commodity in 2026. Claude, GPT, and Gemini all write competent code, and open-weight models like Llama, Mistral, and Qwen run on hardware you control when residency rules under PDPL or DIFC require it. What separates a working system from a stalled pilot is the harness, the evals, and the engineers on your team who own both after the engagement ends.
Visit pooyagolchian.com/contact/ to send a project inquiry, or email hello@pooyagolchian.com directly. Pooya replies within 24 hours. Engagements start with a free consultation to scope requirements and agree on a timeline. He works with startups, scale-ups, and enterprises in Dubai, the GCC, and globally.
Newsletter
Weekly field notes on agentic AI, automation, and high-performance Next.js builds. Each edition is concise, implementation-ready, and tested in production work.
Open full subscription pageGet the latest insights on AI and full-stack development.