Skip to content
Ahmed Taha
عربي
Available for workApplied AI engineer · 19 · Egypt

I build AI systems that run in production.

Three of them are live right now: customer service for online shops, a community for people learning German, and a learning platform with exactly one student — me. I built all three on my own. Every case study also says what still doesn't work, because that's usually the part worth reading.

All workcontact@ahmedtaha.me
Ahmed Taha

This weekFinishing Franco-Arabic language handling in the TalosOps agent.

  • Next.js 15
  • TypeScript strict
  • Supabase / Postgres
  • Row-level security
  • pgvector
  • Gemini · Claude · OpenAI
  • Cohere embeddings
  • MCP
  • Shopify · Paymob
  • WhatsApp · Instagram · Messenger · Telegram · Gmail
  • Aramex · Bosta · DHL · SMSA
  • Notion · Sheets · Drive · Confluence · Miro
  • next-intl (EN/AR)
  • Tailwind v4
  • Vercel
  • Next.js 16
  • React 19
  • TypeScript
  • Gemini
  • Paymob
  • next-intl (DE/EN)
  • Framer Motion
  • react-three-fiber
  • Web Push
  • Supabase / Postgres 17
  • Vitest + embedded Postgres
  • Anthropic / OpenAI / Gemini
  • Vercel (fra1)

Figures

3

Platforms live

511,609

Lines TypeScript

7,878

Tests

586

Migrations

1,017

Commits in 2026

Read from origin/master on 7 Sept 2026, across three repositories. Not estimated.

TalosOps

Work

Three platforms, in production.

For each one: what the problem was, what I built, the decisions I'd defend if you pushed back — and what still doesn't work.

All work
01

TalosOps

One platform running customer service for many businesses at once — an agent that answers and acts across every channel, and Talos, an operator the owner talks to in Egyptian Arabic that runs the platform for them.

In production2026 — ongoingtalosops-ai.com
Next.js 15TypeScript strictSupabase / PostgresRow-level securitypgvectorGemini · Claude · OpenAICohere embeddings
TalosOps product screen

Case study

02

D Plus+

A feed pitched at your level, courses, AI pronunciation practice, full telc/Goethe/ÖSD mock exams with a fresh sheet every time, publish-first moderation, and Egyptian card payments.

Live2026 — ongoingdplus21.com
Next.js 16React 19TypeScriptSupabase / PostgresRow-level securityGeminiPaymob
D Plus+ product screen

Case study

03

nat3alm

A full applied-AI curriculum — 16 phases, 120 lessons, written not generated — with spaced review, a tutor that gives hints before answers, and hand-written Arabic on every single item.

Live2026 — ongoingnat3alm.space
Next.js 16React 19TypeScriptSupabase / Postgres 17Vitest + embedded PostgresAnthropic / OpenAI / GeminiTailwind v4
nat3alm product screen

Case study

Capability

The parts that took the longest.

Each one links to the part of the case study where you can watch it working.

Talos, the operator agent
94 actions, two speeds — approve every change, or let it run and record — and multi-step plans approved once. It runs the platform for the owner, in Egyptian Arabic.
An agent that can act
Answers in the customer's dialect from the business's own knowledge and live data, changes an order only with confirmation and verified identity, and hands off to a human.
Knowledge across languages
A question in English or Franco-Arabic finds the right Arabic document — across Notion, Drive, Sheets, Confluence, Miro, files and page images.
Every channel, every system
WhatsApp, Instagram, Messenger, Telegram, Gmail, web chat; Shopify, Paymob, four carriers, four CRMs, calendars, an OpenAPI connector and a local bridge.
Measured quality
Every closed conversation evaluated, root causes clustered, a simulation harness before real customers, and a read-only quality copilot.
A feed that follows the learner
Bounded signals, decaying interests, publish-first moderation, and exams assembled from a fresh sheet every time.

Writing

Decisions, and what the other option would have cost.

All writing
7 Sept 20267 min

373,000 lines with an agent: the parts that actually hold

Generating code stopped being the hard part a while ago. Everything I do differently now is about the other half: making it cheap to find out whether the code is wrong.

31 Aug 20265 min

If a ranking term is unbounded, your weights are decoration

A feed score with a small weight on engagement can still be almost entirely engagement. Bounding every term is what makes the numbers in the formula mean anything at all.

24 Aug 20266 min

Publish first, review after

Every moderation design starts with the same question, and almost everyone answers it backwards: does review let content up, or take it down?

Video

Me, building on these codebases.

I haven't started yet. Long episodes on YouTube, short cuts on Instagram and TikTok — recorded on the platforms above, including the sessions where the agent confidently ships something broken and I have to catch it.

Videos

Contact

If any of this is useful to you, write to me.

I'm open to applied-AI and agent work, contract or full time.