Skip to work

00 / Index

Applied AI Engineer
Agentic systems & product development

Marrakech, Morocco
31.6295° N, 7.9811° W

Local
Open to work

  • 01Agentic AI systems
  • 02Business workflow automation
  • 03Products from 0 → 1

I turn how a business works into software that works for it. I build agent systems and use coding agents to deliver products, from workflow analysis to implementation.

I learn fast, adapt quickly, and ship quickly on a strong engineering foundation.

Taha Koulal

Scroll — systems under the surface

Strong
foundations.
Learn fast.
Ship fast.

01 / 10 — Multi-agent AI project factory

Factory

An agent-based delivery system where ideas are evaluated, delegated and built by small, specialised teams of AI agents.

Role
Architecture and system design
Stack
Claude Code · MCP · Agent orchestration
Fig. 01 — Delegation graph
Persistent state · audit logIncoming ideaCEO agentevaluate · delegateCTO · mode 1CTO · mode 2CTO · mode 3SandboxFocused build agentsPermission gateMilestone reviewSLA escalationMilestone shipped

01.1 — Context

A CEO agent evaluates incoming ideas, then delegates execution to specialised CTO modes that coordinate focused build agents.

01.2 — Problem

Letting agents execute projects takes more than prompts. Without clear authority, permissions and accountability, execution stops being predictable.

01.3 — What I built

  1. 01The delegation model: CEO agent → CTO modes → focused build agents.
  2. 02Permission gates, milestone reviews and SLA-based blocker escalation.
  3. 03Persistent state, audit logs and sandboxed execution.

01.4 — System

Small, specialised agent teams with explicit responsibilities, persistent state and review gates to keep execution traceable and accountable.

02 / 10 — Autonomous AI app factory

Mobile_factory

A crew of AI agents that finds, validates and builds mobile apps around the clock, while I watch and steer them live from my phone.

Role
Designed and built — agent runtime, factory board, access control, dashboard and hosting
Stack
Claude Agent SDK · Node.js · TypeScript · React · Vite · Server-Sent Events · Oracle Cloud VM · Vercel
Live
sonic-station.vercel.app ↗
Status
In development
Fig. 02 — Factory floor
YouDashboardVercel · React/v1 · HTTPSOracle VM · always onPIN · 8 digits2 misses · lockedSonicplans · assignsScoutStrategistApp builderBackendRelease · QAevery 5 hSettled → SonicPermission policyAuto · full accessKeys · neverMoney → asks youApproval inboxNo answer → denyEvery action → one typed event · SSEA dropped link replays what it missed

02.1 — Context

Claude Code runs headless on a cloud VM, 24/7. A dashboard on Vercel lets me follow and steer the agents from any phone or laptop.

02.2 — Problem

Agents that work alone need real freedom, but some calls still belong to a human. And the work has to keep going when no browser is open.

02.3 — What I built

  1. 01A factory board: a boss agent, Sonic, turns goals into projects and assigns tasks to five specialists: Scout, Strategist, App builder, Backend, and Release & QA.
  2. 02Each agent loads only its own skills and keeps its own notebook. Each task is its own Claude Code job, so it survives restarts and can resume its session.
  3. 03A permission policy: agents have full access to their machine, the server's keys stay out of reach, and anything that spends money goes to me for approval, with a deadline.
  4. 04A Scout that collects App Store charts every five hours without using Claude, then rates up to three ideas, each claim with a source and a date.

02.4 — System

Every agent action becomes one typed event, streamed to the dashboard over Server-Sent Events. A dropped connection gets back exactly what it missed.

02.5 — Access

  1. 01No email or password. A browser is paired once with a one-time code from the server, then opened with an 8-digit PIN.
  2. 02Two wrong PINs lock PIN entry on the server until a single-use emailed link is confirmed. The lock never stops running agents.
  3. 03The browser never sees a token or a key: only HttpOnly, SameSite=Strict cookies, served through the dashboard's own origin.

02.6— Livesonic-station.vercel.app

Open dashboard (opens in a new tab)

03 / 10 — Business operations & technical AI platform

Phova-edge

A platform that digitises PHOVA's departmental workflows and brings technical document extraction, knowledge retrieval and pump recommendations into the same system.

Role
Software engineering intern — workflow analysis, system design and implementation
Stack
Next.js · FastAPI · PostgreSQL · LangChain · ChromaDB · scikit-learn
At
PHOVA Technology
Fig. 03 — Opportunity lifecycle
S.01SalesS.02TechnicalevaluationS.03CostingS.04SourcingS.05DeliveryOpportunityWorkflow rulesNotificationsSalesEngineeringSourcingManagementRole-based accessBefore — multi-step screenAfter — aligned to responsibilitySalesEngineeringSourcing

03.1 — Context

I analysed how PHOVA's departments work, including their responsibilities, information needs and handoffs, to translate day-to-day operations into a digital system.

03.2 — Problem

Digitising operations means understanding the work before designing the software: who owns each step, what information they need and how the next department takes over.

03.3 — What I built

  1. 01Mapped departmental workflows and translated business requirements into Phova-edge's structure and interfaces.
  2. 02Built authentication and user administration with JWT access tokens, refresh cookies and role-based permissions.
  3. 03Implemented core opportunity management and a role-gated state machine with 12 statuses and 18 transitions, alongside notifications, activity history and departmental SLA tracking.
  4. 04Built the UI/UX design system and most frontend pages; implemented KPI computation, with dashboard charts still planned.

03.4 — System

A monorepo with a Next.js App Router frontend and a Python FastAPI backend. Domain modules separate routers, business services and repositories, using async SQLAlchemy with PostgreSQL and Alembic migrations. Docker and Docker Compose run the application services.

03.5 — Technical AI — documents to structured data

  1. 01Built processing paths for PDF, Excel and Word documents using PyPDF2, pdfplumber, Docling, pandas, openpyxl and python-docx; scanned pages use pdf2image, Poppler and configurable vision providers.
  2. 02Extracted pump tags, flow, head, fluid properties and component materials into structured JSON, normalising French and English terminology through a custom technical dictionary.
  3. 03Used EasyOCR and OpenCV to locate extracted values on page images for review and highlighting.

03.6 — Retrieval & recommendations

  1. 01Built RAG retrieval with LangChain, ChromaDB and Sentence Transformers (all-MiniLM-L6-v2) to support extraction and material/foundry recommendations.
  2. 02Implemented Random Forest classifiers for pump types and references with scikit-learn, pandas and NumPy, with SHAP and language models for prediction explanations. Joblib stores models, encoders and feature definitions.
  3. 03Added custom Python hydraulic checks and configurable provider routing with fallback rules for Groq, Gemini and optional DeepSeek. Active providers and models depend on deployment configuration.
  4. 04Built document chat using a Groq text model and custom prompts, passing up to 14,000 characters directly as context; this chat path is separate from retrieval.

04 / 10 — Football identity mobile product

Vamo

A mobile product that gives football players a persistent identity across matches, teams and tournaments — including a Football Passport.

Role
Designed and built — product structure, core flows, data access
Stack
React Native · Supabase · Row-level security
Live
vamo-l.vercel.app ↗
Fig. 04 — Identity graph
Team 1Team 2MatchAttendanceTournamentPhone · teammates onlyRLS policyPlayerFootball Passport

04.1 — Context

Players, teams, matches and tournaments form a dense web of relationships. Vamo turns it into a persistent profile history.

04.2 — Problem

The network has to be discoverable, while personal data stays visible only to the right people.

04.3 — What I built

  1. 01Product structure and core flows from zero: discovery, match and tournament creation, attendance, profile history.
  2. 02Persistent player identities, teams, tournaments and the Football Passport.
  3. 03Privacy-aware Supabase RLS, including teammate-only phone visibility and secure player-data access.

04.4 — System

React Native client on Supabase. Row-level security policies decide who can read which player data.

04.5— Livevamo-l.vercel.app

Visit landing page (opens in a new tab)

05 / 10 — Real-time multiplayer platform

Transcendence

A multiplayer Pong platform with authentication, live messaging and synchronised game state.

Role
Auth, live messaging, real-time game state and session behaviour
Stack
React · TypeScript · WebSockets
Fig. 05 — Event stream
Client AServerClient BClient AWebSocket srvClient Bauthsessioninputgame statechatchatsyncsync20+ concurrent games

05.1 — Context

Game state, chat and user interactions change at the same time, for many players, across many games.

05.2 — Problem

Every client has to agree on what is happening — the ball, the score, the conversation — while events keep arriving.

05.3 — What I built

  1. 01Authentication, live messaging and synchronised game state.
  2. 02Real-time frontend state and game-session behaviour supporting more than 20 concurrent games.
  3. 03Event-driven communication to keep multiplayer state, chat and interactions in sync.

06 / 10 — AI wardrobe and outfit app

FIT

Catalogue the clothes you already own once. Get a complete, context-aware look each day — instead of another purchase.

Role
Product concept and recommendation structure
Stack
React Native · Recommendation system
Status
Product concept
Fig. 06 — Recommendation model
Wardrobe itemsWeatherSkin toneBody shapeColour compatibilityPreferencesContextComplete lookReuse, not purchase

06.1 — Signals

Wardrobe items, weather, skin tone, body shape, colour compatibility and personal preferences.

06.2 — Output

Daily outfit recommendations built from wardrobe reuse rather than unnecessary purchases.

07 / 10 — Unix shell written in C

Minishell

A working shell: it parses a line the way bash does, then runs it with fork, pipe, dup2 and execve — without leaking a descriptor.

Role
Parsing, execution and signal handling
Stack
C · Unix processes · Signals
At
42 Network (1337)
Fig. 07 — Minishell, running
White hedgehog pixel art — Sonic, my nickname

Sonic — learn fast. Build solid. Ship fast.

minishell 1.0 · 42 network

Same grammar and builtins as the shell in C, modelled in the browser.

Type `help`, or `ls` to read this site as files.

 

minishell:~$ cat about.txt

Taha Koulal — Applied AI Engineer

Agentic systems & product development

 

Sonic is my nickname: I learn fast, adapt quickly,

and ship quickly on a strong engineering foundation.

 

location Marrakech, Morocco (31.6295° N, 7.9811° W)

email tkoulal.tech@gmail.com

github https://github.com/tahariginal

 

Available for full-time and early-stage product opportunities.

minishell:~$

07.1 — Context

Everything a shell receives is one flat string. Before anything can run, it has to become tokens — words, operators, and quoted spans that mean themselves.

07.2 — Problem

Quotes, variable expansion, pipes and redirections all act on the same line, and the order they are resolved in decides what actually runs.

07.3 — What I built

  1. 01A tokeniser that respects single and double quotes, then a parser that turns the line into a pipeline of commands with their redirections.
  2. 02Expansion of $VAR and $? after tokenising, never inside single quotes.
  3. 03Execution with fork, pipe, dup2 and execve, PATH resolution, and heredocs — with builtins run in the parent when they change its state.
  4. 04The builtins the subject asks for: echo -n, cd, pwd, export, unset, env, exit.

07.4 — System

Interactive signal handling: Ctrl-C returns a fresh prompt and status 130, Ctrl-D ends the session, and every descriptor opened for a redirection is closed. The figure is that shell, modelled in the browser — type into it.

08 / 10 — Playable 3D grand piano

Piano

A grand piano you can play in the browser, with real Yamaha C5 recordings, guided lessons, melodies and a recorder.

Role
Designed and built — audio engine, 3D instrument, input, lessons and recorder
Stack
Next.js · React · TypeScript · three.js · Web Audio API · Tailwind CSS · Vercel
Live
piano-sand.vercel.app ↗
Fig. 08 — Signal path
Ode to Joy · 104 BPMOctave 412345678C4C5Shift · black keys[ ] octave · Space pedalKeysTouchControllerHolds · counts2 holders · 1 noteSounds once1 · Sound first2 · Listeners21 recordings · C2–C7ReverbSampleLowpassGainBusLimiterOutWeb Audio · no audio librarySceneDraws on changeLessonsNotes · orderRecorder→ TimelinePlayerReplay · demo120 ms ahead

08.1 — Context

Number keys 1–8 play an octave, Shift adds the black keys, [ and ] change octave and Space holds the pedal. A mouse or a finger plays too, and slides across the keys.

08.2 — Problem

The keyboard, a finger, a replay and a lesson demo can all press the same key at once. The note has to sound the moment it is struck — once, and without waiting for React or the next frame.

08.3 — What I built

  1. 01A sampler on the Web Audio API, with no audio library: 21 Yamaha C5 recordings from C2 to C7, one every minor third, so every key is at most a semitone from a real recording.
  2. 02Each voice has its own filter and release, so soft strikes sound darker. Playback starts on the hammer, not on the silence at the start of the file.
  3. 03A sustain pedal, a soft reverb and a limiter. Notes fire straight from the input handler, before anything is drawn.
  4. 04Keys are read by position, not by character, so AZERTY, QWERTZ and Arabic layouts play the same notes.

08.4 — Instrument

A small grand modelled in code with three.js: notched keys on hidden pivots, strings drawn as hairlines, an open lid on its prop. It draws a frame only while something moves, so an idle piano costs nothing.

08.5 — System

  1. 01One controller holds every note, under the name of whatever pressed it. A note sounds while at least one holder holds it, so two inputs never strike it twice.
  2. 02Takes, melody demos and lesson demos become one list of timed events. The player schedules notes 120 ms ahead on the audio clock and moves the real keys.
  3. 03Six guided lessons and five melodies check which notes you play and in what order, not rhythm. The recorder keeps timing, chords and the pedal.

08.6— Livepiano-sand.vercel.app

Play the piano (opens in a new tab)

09 / 10

NDA · Confidential

Confidential project 01

Worked on a project covered by a non-disclosure agreement. Client identity, technical details and deliverables are confidential.

10 / 10

NDA · Confidential

Confidential project 02

Worked on a second project covered by a non-disclosure agreement. Client identity, technical details and deliverables are confidential.

02 / Experience

From the surface down.

Production work on top. The interactive and systems foundations it stands on, underneath.

  1. L.01 — Surface

    PHOVA Technology

    Software Engineering Intern

    Production

    Built Phova-edge to digitise business operations, starting by analysing how sales, engineering, sourcing and management work and hand information between departments.

    Owned authentication, user administration, core opportunity workflows, notifications and UI/UX across Next.js and a layered FastAPI/SQLAlchemy backend on PostgreSQL. Also implemented KPI computation, with dashboard charts planned.

    Built technical AI workflows for document extraction, RAG retrieval and pump/material recommendations, combining configurable language and vision providers with Random Forest predictions, SHAP explanations and hydraulic calculations.

    • Business process analysis
    • System design
    • Workflow digitisation
    • RAG
    • Document AI
  2. L.02 — Interactive

    UM6P — IDC Morocco

    VR Development Intern

    VR / Interactive Technologies

    Developed an immersive virtual-visit experience from prototype to functional delivery, including interactive 3D environments and navigation.

    • Unity
    • C#
    • 3D navigation
    Fig. L.02 — Virtual visit, stereo model
    ABC
    Room
    A · Entrance
    Pos
    0.0, -4.5
    Head
    —
    Frame
    —

    Waypoint 01 / 04

    Model, not a capture · drag or arrow keys to look · click a ring or press 1–4 to jump · gaze at ◇ to open info
  3. L.03 — Foundation

    42 Network (1337)

    Common Core — in progress

    Software Engineering

    Validated ft_irc, Minishell, Philosophers, ft_printf, Libft and 10+ projects — systems programming from the ground up.

    • ft_irc
    • Minishell
    • Philosophers
    • ft_printf
    • Libft

03 / System

The stack, in cross-section.

Grouped by where each piece sits relative to the surface. Select a technology to trace it back to the work.

01  Product engineering

Above the surface — what people touch

02  Backend & systems

Under the surface — state, data, transport

03  Applied AI & agent systems

Agents, document understanding, retrieval and prediction

Also — Unity · C# (VR internship, UM6P)

Method

How I approach a build.

  1. i

    Understand the business

    Learn how teams work, map their responsibilities and handoffs, and identify what slows them down.

  2. ii

    Design for growth

    Turn operational needs into clear workflows, data models and system boundaries, planning for scale and maintainability.

  3. iii

    Engineer for reliability

    Choose algorithms and workflow logic for the problem, with clear permissions, failure handling and checks at critical boundaries.

  4. iv

    Build, check, adapt

    Build agent systems and use coding agents to deliver products. Learn the tools each problem needs, verify results and refine with user feedback.

Engineering strengths

  • 01Business process analysis & digitisation
  • 02Agent systems & tool integration
  • 03System design for scalability & reliability
  • 04Product delivery with coding agents
  • 05Rapid learning & adaptation
  • 06RBAC, data boundaries & review gates
Portrait of Taha Koulal, smiling, wearing glasses and a black hoodie
Photo style

The original photo. A little room to experiment below.

04 / About

Understand the work. Build the system.

I build agentic AI systems and use coding agents to turn ideas into working products. I am effective & efficient in how I work. I learn unfamiliar tools and domains quickly, starting with the problem and the people doing the work.

At PHOVA, I analysed how departments operate and built Phova-edge to digitise their workflows. That means translating business needs into system design, planning for growth and reliability, and making everyday tasks simpler.

My foundation is C, Unix and networking at 1337, alongside interactive development at UM6P. I bring that engineering perspective to AI: clear permissions, deliberate technical choices and review of what agents produce.

Based in
Marrakech, Morocco
Studying
42 Network (1337)
Languages
Arabic — Native English — Professional French — Intermediate
Open to
Full-time & early-stage product roles

05 / ContactAvailable for full-time and early-stage product opportunities

Have something unclear? Let's make it work.

Write to