Eslam Hamed
Technical Lead & AI Solutions Architect
Building enterprise AI systems, cloud platforms, and production-ready software
Writing
I write about production AI the way I build it: what the constraint was, which decision mattered, what it cost, and the numbers afterwards. Most of it comes from shipping retrieval and agent systems on AWS and Azure for enterprise customers, and from building products alone. A few posts are experiments from outside work; they sit in their own section.
Start here
Four guides on how I build production AI: securing agents, answers you can verify, evaluation before launch, and cost without quality loss.
-
· 7 min read · AISecurity
Designing Secure AI Agents for Enterprise Data
One prompt injection in a supplier PDF, followed through eight controls for AI agents on enterprise data: what each stops, what it misses, how to test it.
-
· 6 min read · AgenticAI
Building a Governed AI Analytics System with Verifiable Answers
How to build an AI analyst executives can trust: the model plans, a deterministic engine computes, and every figure is inserted from a receipt the model cannot write.
-
· 6 min read · RAG
How to Evaluate a Production RAG System Before Deployment
A practical way to evaluate a RAG system before customers use it: build the test set, score retrieval and answers apart, test refusal, and set a release gate.
-
· 5 min read · LLMOps
How to Reduce LLM Costs Without Regressing Answer Quality
How I would cut LLM spend safely: find where the money goes, change one thing, prove quality held on a frozen suite, and roll out with a way back.
From the work
-
· 7 min read · LLMOps
Claude Fable 5.2 and Mythos 5.2: Benchmarks, Pricing, and When to Choose Them Over Opus 5.5
Claude Fable 5.2 leads Anthropic's coding table and costs 20% less than Fable 5.1, yet twice Opus 5.5 per token. What the numbers mean, and when to pick it.
-
· 8 min read · AISecurity
GPT-6 Cyber vs Kimi K3.1 Pro: Gate the Hacking Capability, or Remove It?
OpenAI released GPT-6 Cyber only to vetted security teams. Moonshot removed cyber data from Kimi K3.1 Pro before opening its weights. What each approach buys.
-
· 9 min read · OpenWeights
GPT-6 Astra Repaired Qwen3.5-0.8B by Editing Its Weights: What the MIT Result Really Shows
MIT researchers had GPT-6 Astra change 0.016% of Qwen3.5-0.8B's weights, and MATH-500 rose from 7.9% to 48.1%. Why it worked, what it cost, and what it cannot do.
-
· 6 min read · SoftwareEngineering
Stop Rebuilding the Same Things: How I Built a Personal Engineering Toolkit
Every project starts by rebuilding sign-in, evals and document parsing. How I turned my past work into a reusable toolkit.
-
· 6 min read · LLMOps
Gemini 4 Carbon vs Claude Opus 5.5: Benchmarks, Cost, and Production Readiness
Gemini 4 Carbon vs Claude Opus 5.5: Carbon leads the Artificial Analysis index by one point, Opus leads on coding agents. What that means for cost and production.
-
· 10 min read · MCP
MCP Implementation Roadmap: A Six-Phase Enterprise Rollout Plan for the Model Context Protocol
An MCP implementation roadmap for companies: six phases from a read-only pilot to approved writes, each with an exit test, and what the 2026-07-28 spec requires.
-
· 8 min read · AISecurity
LiteLLM Compromised Twice: Why a Valid Signature Did Not Help, and How to Protect Your AI Gateway
LiteLLM's v1.107.0 was built by its own pipeline from a lookalike fork, then signed. What happened in March and October, why signatures failed, and a gateway checklist.
-
· 8 min read · RAG
Cohere Rerank 5: Better Retrieval, Less Context, and What the Benchmarks Really Show
Cohere Rerank 5 adds Pro and Fast models and Context Focus, which picks the passages that matter inside each document. What the benchmarks show, and how to test it.
-
· 8 min read · OpenWeights
Meta Muse Spark Open: Licence Terms, the Hardware Bill, and What Its Safety Training Promises
Meta released Muse Spark Open's weights on October 15. What the Muse Model License requires, what 816B parameters cost to host, and what safety training can promise.
-
· 10 min read · LLMOps
Claude Haiku 5.5: Benchmarks, Pricing, and Why Cheaper Tokens Don't Always Mean Cheaper AI
In a well-designed AI application, changing the model should be an experiment with measurable acceptance criteria, not a rewrite of the system.
-
· 8 min read · LLMOps
GPT-6.1 Sol vs GPT-6 Astra: Benchmarks, Pricing, and the Real Cost of AI Agents
GPT-6.1 Sol vs GPT-6 Astra: one point apart on the Artificial Analysis index, at a fifth of Astra's token price. Whether that saving survives real agent workloads.
-
· 8 min read · AICoding
Claude Opus 5.5 vs GPT-6 Astra: Coding Benchmarks, Cost, and What Matters in Real Projects
A capable coding model can write a large amount of code. A dependable engineering process determines what is safe to keep.
-
· 6 min read · AISecurity
Google's Gemini Reached Three Real Companies in a Test: What Is Known and What Is Not
Google confirms a Gemini model accessed three real companies during a May test at Irregular. What is confirmed, what is missing, and why disclosure matters.
-
· 10 min read · AgenticAI
Claude API On-Demand Compaction: Solving Long-Running Agent Context Without Losing Critical Information
A model-generated summary is useful for continuity. Reliable execution depends on knowing which information can be summarized and which must remain verifiable.
-
· 6 min read · LLMOps
DeepSeek V4.1 Flash: Coding Benchmarks, Asymmetric Architecture, and the Economics of AI Agents
The question is whether V4.1 Flash completes the same work, to the same acceptance standard, at a lower total cost.
-
· 10 min read · RAG
How LLMs Like ChatGPT or Gemini Turn Customer Feedback into Business Recommendations
ChatGPT and Gemini do not learn from your customer feedback on their own. How it reaches the model, the pipeline behind business recommendations, and where it fails.
-
· 6 min read · AISecurity
GPT-6 Astra: Frontier Benchmarks, Critical Cybersecurity, and the Risk of Autonomous AI Agents
A smarter model can make an agent more useful. It should never become the sole authority deciding what the agent is permitted to do.
-
· 6 min read · AgenticAI
Building a Governed AI Analytics System with Verifiable Answers
How to build an AI analyst executives can trust: the model plans, a deterministic engine computes, and every figure is inserted from a receipt the model cannot write.
-
· 6 min read · AgenticAI
Anthropic's Model Hardware Standard: When AI Agents Start Controlling Physical Machines
Anthropic's Model Hardware Standard lets AI agents operate lab and factory equipment. How MHS relates to MCP, and why safety limits belong in the device controller.
-
· 7 min read · AgenticAI
1,200 Agents and a Message Board: What METR and OpenAI Found in the Hugging Face Incident
METR and OpenAI's reports show agents built a covert message board, recruited each other and attacked Hugging Face to beat a grader that never checked.
-
· 8 min read · MCP
MCP's August 2026 Roadmap: Stateless Servers, Agent Identity, and Enterprise Security
MCP's July 28 specification made the protocol stateless; the August 22 roadmap puts agent identity and tool discovery next. What shipped, what is planned, how to prepare.
-
· 6 min read · OpenWeights
Qwen3.8-2.4T-A95B: Benchmarks, Architecture, and the Real Hardware Bill for Self-Hosting
Qwen released 2.4T open weights on August 12. What the benchmarks cover, why a 4-bit checkpoint is 1.45 TB, and how much hardware it really takes to serve.
-
· 6 min read · OnDeviceAI
Liquid AI LFM2.5-2.6B: What a 2.7B-Parameter Agent Model Can Really Do on a Device
Liquid AI's LFM2.5-2.6B targets tool calling on laptops and phones. What the numbers show, where it trails, and what the licence means for commercial use.
-
· 7 min read · AISecurity
AI Sandbox Escape Reports: What Anthropic, UK AISI, OpenAI and Meta Actually Said
After OpenAI's breach, Anthropic, UK AISI, OpenAI and Meta reported agents reaching real systems. Most were open internet links, not escapes. Why that matters.
-
· 7 min read · AISecurity
Designing Secure AI Agents for Enterprise Data
One prompt injection in a supplier PDF, followed through eight controls for AI agents on enterprise data: what each stops, what it misses, how to test it.
-
· 6 min read · OpenWeights
Kimi K3 Open Weights: Benchmarks, Licence, and What It Takes to Run 2.8T Parameters
Kimi K3's open weights landed on July 27: 2.8T parameters, 104B active. Independent scores, the real licence terms, and why the 1.56 TB download is the easy part.
-
· 6 min read · AISecurity
OpenAI and Hugging Face: How Models Under Test Escaped a Sandbox and Breached Production
GPT-5.6 Sol and a research model left a test sandbox through a package proxy zero-day and reached Hugging Face. What failed, what held, and what to change.
-
· 6 min read · AISecurity
AI Coding Agent Sandbox Escapes: How Cursor, Codex, Gemini CLI and Antigravity Were Bypassed
Pillar Security found seven ways coding agents got out without breaking the sandbox: they wrote files that trusted tools later ran. The pattern and the fixes.
-
· 5 min read · AISecurity
The Hugging Face Intrusion: An AI Agent Attacker, and Guardrails That Blocked the Defenders
Hugging Face says an autonomous agent ran its July intrusion, and that hosted AI guardrails refused its forensics. What defenders should prepare in advance.
-
· 6 min read · AISecurity
Grok Build Uploaded Whole Repositories: What the Wire Tests Show About AI Coding CLIs
A proxy capture showed Grok Build sending entire git repositories to remote storage, and a training opt-out did not stop it. What was measured, and how to test.
-
· 6 min read · DotNet
.NET 8 and .NET 9 End of Support: Move to .NET 10 Now or Wait for .NET 11?
.NET 8 and .NET 9 both end support on November 10, 2026. What breaks on the way to .NET 10, why waiting for .NET 11 is risky, and how to test the upgrade.
-
· 5 min read · LLMOps
How to Reduce LLM Costs Without Regressing Answer Quality
How I would cut LLM spend safely: find where the money goes, change one thing, prove quality held on a frozen suite, and roll out with a way back.
-
· 6 min read · LLMOps
Claude Fable 5 Suspension: What a 19-Day Model Outage Means for Production AI
Export controls took Claude Fable 5 offline for every user three days after launch. What the 19-day gap teaches teams that depend on one frontier model.
-
· 8 min read · OpenWeights
GLM-5.2 Open Weights: Benchmarks, MIT Licence, and the Hardware to Self-Host a 1.5 TB Model
Z.ai released GLM-5.2 under MIT with a 1M-token context. What the vendor benchmarks show, what they leave out, and the memory arithmetic for running it yourself.
-
· 6 min read · RAG
How to Evaluate a Production RAG System Before Deployment
A practical way to evaluate a RAG system before customers use it: build the test set, score retrieval and answers apart, test refusal, and set a release gate.
-
· 6 min read · ModelEvaluation
DeepSWE vs SWE-Bench Pro: What a 24% False Negative Estimate Means for Coding Benchmarks
Datacurve's DeepSWE audit estimates SWE-Bench Pro verifiers wrongly rejected 24% of sampled trials. How to read coding agent leaderboards and test your own.
-
· 8 min read · AgenticAI
Claude Managed Agents, Gemini and Microsoft Agent Framework: Build or Buy Your Agent Runtime?
Anthropic, Google and Microsoft now host your agent sessions. How isolation, state, pricing and lock-in compare, and what Agent Framework 1.0 means for Semantic Kernel.
-
· 6 min read · AgenticAI
PocketOS, Cursor and Railway: What the Nine-Second Database Deletion Really Teaches
PocketOS's founder says a Cursor agent on Claude Opus 4.6 deleted a production volume with one API call. What Railway confirms, and which controls would hold.
-
· 7 min read · LLMOps
GPT-5.5 vs Claude Opus 4.7: API Pricing, Tokenizers, and the Real Cost per Task
GPT-5.5 charges more per output token than Claude Opus 4.7, yet matched its top index score for a quarter of the cost. Why cost per task is the figure to compare.
-
· 6 min read · AISecurity
Claude Mythos Preview Sandbox Escape: What Anthropic's System Card Actually Says
Anthropic's Mythos Preview card describes a sandbox escape it asked for and a public post nobody requested. What the coverage missed, and what to change.
-
· 8 min read · OpenWeights
DeepCoder Explained: Microsoft's 2017 Research vs Agentica's DeepCoder-14B, Worth Running in 2026?
There are two DeepCoders: Microsoft Research's 2017 program-synthesis paper and Agentica's open-source DeepCoder-14B. What each one is, and whether to run the 14B now.
-
· 8 min read · AICoding
Claude Code Source Map Leak: How a .map File Reached npm, and the Release Checks That Stop It
Claude Code 2.1.88 reached npm with a 59.8 MB source map. How map files end up in packages, and the release checks that stop them: allowlists, dry runs and diffs.
-
· 6 min read · AISecurity
Alibaba-Linked ROME Agent and the Crypto Mining Story: What the Paper Actually Says
An Alibaba-affiliated paper says an agent opened a reverse SSH tunnel and diverted GPUs to mining in training. What it shows, and what the headlines added.
-
· 8 min read · AgenticAI
GPT-5.4 Tool Search and Long-Context Pricing: What the 272K Line Means for Agents
GPT-5.4 can defer MCP tool schemas with tool search and read about 1.05M tokens, but prompts over 272K cost 2x input. What that means for agent design and cost.
-
· 12 min read · SoftwareEngineering
The End of Specialization: Why the Software Industry Is Entering a New Era of Engineering
Front-end, back-end, QA and analyst roles are converging as AI tools make routine code cheap. Why systems thinking and engineering judgement now outlast syntax skills.
-
· 4 min read · SEO
Today I Used AI to Make My Portfolio Easier to Find on Google
How AI Helped Me Improve Portfolio SEO and Performance
-
· 8 min read · AICoding
Claude Opus 4.6 Agent Teams Wrote a C Compiler: What $20,000 Bought and What It Did Not
Sixteen Claude Opus 4.6 agents spent two weeks and nearly $20,000 writing a C compiler that builds Linux. What the tests specified, and what still needed a human.
-
· 8 min read · AISecurity
OpenClaw Gateway Takeover, Moltbook's Open Database and Malicious ClawHub Skills: What to Change
A one-click OpenClaw gateway takeover, a Moltbook database without row-level security and 341 malicious ClawHub skills: what failed, and what to change.
-
· 8 min read · MCP
MCP Apps, the First Official MCP Extension: Interactive UIs, Sandboxing and What to Test
MCP Apps lets MCP tools return interactive UIs that render in sandboxed iframes in Claude, VS Code and Goose. How it works, where the risks are, and what to test.
-
· 5 min read · Metrics
When the Metric Becomes the Target: Goodhart's Law in Software
Why tickets closed, story points and test coverage get gamed once they become targets, what Wells Fargo shows, and how counter-metrics keep a target honest.
-
· 5 min read · Product
Every Custom Feature Comes With Interest
Why customer-specific features cost more every release, how configuration, customisation and a fork differ, and how to price custom work for its whole life.
-
· 5 min read · Architecture
Conway's Law: Your System Copies Your Org Chart, and Your Client's
What Melvin Conway wrote in 1968, why a large client's departments reshape your integration, and the questions to ask before you draw any service boundaries.
-
· 5 min read · Customers
The Customer Who Complains Is Still Talking to You
A complaint is information and silence is exit. What the peak-and-end studies suggest about handling one, and how to investigate it without letting it set the roadmap.
-
· 5 min read · Contracts
Fixed Price or Time and Materials: Who Carries the Risk?
How fixed price and time and materials split the uncertainty in a software contract, what each one rewards, and the middle options that make the risk explicit.
-
· 5 min read · Prioritisation
Cost of Delay: Put a Price on Waiting
Rank work by what each week of waiting costs, divided by how long it takes: cost of delay and CD3, a worked example, and questions for planning.
-
· 5 min read · BuildVsBuy
Build or Buy? Either Way, You Pay for Something
Build or buy, compared over the life of the system: the costs a two-month estimate leaves out, why teams lean towards building, and a decision table to reuse.
-
· 5 min read · Leadership
Your Best Engineer Might Be Your Biggest Risk
Why the engineer who knows everything becomes a queue and a risk, what the bus factor measures, and how to spread knowledge before the next holiday tests it.
-
· 5 min read · TechDebt
Technical Debt Is a Loan: Explain It With Interest, Not Code
Explain technical debt to the people who pay for it: principal, interest in days per change, and why to repay where code changes most, not where it looks worst.
-
· 5 min read · Product
The Feature Nobody Used, and What the Kano Model Says About It
A feature can pass every acceptance test and change nothing. What the Kano model's paired questions reveal about must-be, attractive and indifferent features.
-
· 5 min read · Estimation
"How Long Will It Take?" Why a Software Estimate Is Not a Promise
The decision behind 'how long will it take?', how estimates, targets and commitments differ, and how to answer with ranges, assumptions and scope that can move.
-
· 5 min read · Scope
"It's Just a Small Change": The Most Expensive Sentence in Software
One extra field followed through migration, reports, APIs, integrations and support, with a ripple checklist and a way to state the cost without sounding defensive.
-
· 5 min read · Requirements
Everyone Agreed in the Meeting. Nobody Meant the Same Thing.
Words like fast, simple and secure leave a meeting with three meanings. How to turn each one into a past example and a testable acceptance criterion.
-
· 2 min read · JTBD
Beyond the Obvious: Understanding True Needs
The short film "It's Not About The Nail" serves as a poignant metaphor for understanding others, a scenario that resonates deeply in customer service and interpersonal…
-
· 2 min read · ArtificialIntelligence
Unraveling the Power of Large Language Models in Customer Review Analysis
Decoding Customer Sentiments with AI
-
· 1 min read · ProductDevelopment
"Show Me the Money!" - Understanding the True Purpose of Your Code
It's crucial to understand not just the code you're writing or the feature you're adding, but what it actually does – not for you, but for the customer. The client pays for or…
-
· 1 min read · AI
Is DeepCoder the End of Human Programmers?
For those who are yet unacquainted, DeepCoder is an emerging tool that leans on artificial intelligence to solve programming-related problems. In essence, it analyzes the input…
Off the clock
-
· 3 min read · Troubleshooting
Today I Used AI to Save Movie Night (Yes, Really)
How AI Helped Me Fix a Black-Screen HDMI Issue Fast
-
· 2 min read · HealthSystems
For Three Months, I Used AI to Drop 16 KG (ChatGPT Edition)
How AI Helped Me Build a Fat-Loss System That Actually Fits Real Life
-
· 1 min read · PersonalDevelopment
Destructive Traits: How Your Behavior Affects Those Around You
Inspired by what I've read and resonated with, this article delves into behaviors that can unknowingly harm our relationships. As Cathy Caprino points out in her insightful…
-
· 1 min read · SelfImprovement
Debug Yourself: How Lazy Are You?
Redefining Laziness: A Journey to Self-Motivation
