I build AI products on my own, from the first idea to the day they go live, and then I keep them running.
Most of what I ship is the unglamorous half of AI: the retrieval that has to find the right chunk, the agent that has to refuse when it does not know, the publish gate that stops a model outage from writing nonsense to a customer's Google profile. I have learned most of what I know from things that broke in production rather than from documentation.
I write about that too, including the times I got it wrong.
- Now — agentic SEO automation, autonomous review replies, and an open benchmark for agent memory systems
- Stack — Next.js, React Native, Node, Python, Postgres, MySQL, vector search, LangGraph, cloud
- Based in Pakistan, working with teams in Sweden, the US, the Philippines and the UK
- Open to contract work, and to being sent a system you want measured
|
An honest benchmark for AI agent memory: Mem0, Zep, Letta, LangMem and GoodMem on the same 419-turn conversation, graded only on what each one retrieves rather than what it can generate. Vendors publish 80-93% on LoCoMo. Measured under one fixed, published method, the best scored 57.6%. Every number recomputes from the per-probe log in the repo. It also includes 28 adversarial probes where refusing is the correct answer, because a memory system that always says something confidently is a liability.
|
One conversation, 120 probes, |
| Product | What it does | Live |
|---|---|---|
| Jorge Castro AI | Agentic content and SEO automation, brand visibility tracking across LLMs | jorgecastro.ai |
| Responsly | Autonomous Google review replies in the business's own voice and the customer's language | responsly.ai |
| Skimming | Multi-tenant RAG workspace for bulk video and document understanding | skimming.ai |
| FindUrLawyer | Legal marketplace for the Philippines, matching clients to lawyers | findurlawyer.com |
| Saleslights | AI sales video platform with avatar generation | saleslights.com |
| Video Funker | AI sales video SaaS | videofunker.ai |
| Soor Exchange | Shariah-compliant crypto trading app, React Native | Play Store |
| Outreach Engine | Cold email at scale: sourcing, AI-written openers, 30 mailboxes, reply classification | private |
Seventeen shipped products in total. The rest are on hamzashabbir.dev/work.
|
AI in your product
|
Growth systems
|
Product engineering
|
Plus the ones without a nice icon: LangGraph, LangChain, Qdrant, pgvector, Playwright, Cloud Tasks, Smartlead, and more model APIs than I would like to admit.
I publish what I learn from production, including the failures.
- I benchmarked 5 AI agent memory systems. The best scored 57.6%. — including a correction to a benchmark I previously published without running
- Our LLM fallback worked perfectly. That was the problem. — a fallback that published generic English replies to Swedish businesses for days while every dashboard stayed green
More at hamzashabbir.dev/article.
I take on a small number of contracts at a time. If you are adding AI to a product, or choosing between memory and retrieval systems with real money on the line, I am easy to reach.
And if agent-memory-bench is useful to you, a star helps the next person find a number they can check rather than one they have to take on faith.