Parse 01
An LLM turns messy Excel, CSV, or PDF into one validated JSON contract. It reads the input — it never decides the schedule.
01Build 01 / 08
DWG P-01 · EdTech · scheduling · supervised pilot
SANAD — a timetable that certifies itself.
A bilingual, Arabic-first exam-scheduling system for Al-Ahliyya Amman University. A messy registrar file goes in; a provably conflict-free exam schedule comes out. The model only reads — a deterministic solver builds the grid, and a separate verifier proves zero student conflicts before a single result is shown.
Solve 02
A deterministic engine — graph-colouring (DSATUR) and OR-Tools CP-SAT — builds the timetable. Reproducible, no model in the decision.
Certify 03
A separate verifier re-derives every conflict from the raw enrolments. No certificate of zero conflicts, no render — ever.
0Student conflicts
100%Independently certified
AR/ENBilingual · full RTL
2-layerSolver + verifier
The tour · 01 — the console
One console: detect → build → certify.
Upload a registrar file and the system walks three gates — detect the conflicts, build the grid, certify it — then unlocks every sheet, proof, and export from a single screen.
The tour · 02 — the certified output
The certified timetable.
The final output: every course placed, with its slot, room, seat count, and invigilating instructor. Print-ready, bilingual, and stamped with an integrity fingerprint — shown only because the verifier certified it.
The tour · 03 — the intelligence layer
A situation room for the exam season.
A live regional-risk globe tracks the countries AAU's students come from, flags disruption, and spells out the academic impact and the recommended action — so scheduling sees the whole region, not just the calendar.
An AI aide per professor — grounded, not guessing.
Each assistant is distilled from a faculty member's public expertise. It answers from that profile and openly says when it doesn't know — an aid, explicitly never a stand-in for the professor.
The tour · 04 — the proof it shows its work
The accreditation certificate.
A sealed, zero-conflict certificate — students, professors, and rooms checked — with each guarantee written as the exact rule it proves.
Why the constraints exist.
For every course, the exams it cannot share a slot with — and the precise number of students they have in common. The conflict graph, made legible.
The math, auditable.
Every figure the system reports is backed by its explicit formula, its real inputs, and the result — so staff can check the arithmetic, not just trust it.
Read-only analytics.
Exams, days, seats, rooms, and a student daily-load fairness signal — pure projections over the certified schedule, never a re-solve.
The tour · 05 — the sheets & the toolkit
The course table.
Every scheduled course with code, faculty, date, time, room, and seat count — the machine-readable spine of the schedule.
Proctor distribution.
Each exam with its assigned proctors and supervisor, signature lines and all — the sheet the invigilation office actually prints.
One toolkit for the faculty.
Room availability, exam lookup, workload, deadlines, accommodations, and a live cost calculator — the day-to-day utilities staff reach for, in one tab strip.
02Build 02 / 08
DWG P-02 · Network engineering · solution architecture · engineering prototype
FABRIC — designs that must survive their own critic.
It started as a challenge. A solution-architecture expert at Cisco asked: “tell me what you can build that solves the problems in my work.” His CV became the blueprint — the stages he walks through to solve a network problem, turned into an agent system. State a problem and an orchestrator routes it across 16 specialist agents — discovery to design to config to migration — with hard gates in the path. A design that fails its red-team review, or a config that fails the linter, is routed back, not shipped.
Decompose 01
The orchestrator plans a graph of specialists per problem — Design, Implement, Sell, Operate, Scale — a road network with a driver, not a fixed pipeline.
Gate 02
A Critic red-teams every design before it is shown, and a deterministic linter validates every config. Failing work is sent back with the findings — never presented.
Ground 03
A citation-guard checks every RFC claim against a grounded standards index. Invented citations are blocked, and every run ends with a trust report that says so.
16Specialist agents
13Pipeline stages
300Benchmark questions
47Test batteries · all green
The tour · 01 — the console
State the problem. Watch it argue.
One console: type a network problem and the orchestrator decomposes it across five phases and thirteen stages — every card a specialist agent, every output badged for how it was produced, every stage grounded before the next begins.
The tour · 02 — the gate that catches the lie
Every run ends in a confession.
The trust report is a deterministic gate, not prose: citations verified against a grounded index, a fabricated RFC caught and blocked, the config lint result, and every assumption the system made — listed for the human architect to confirm before anything ships.
The tour · 03 — the working surfaces
Six engagements, ready to run.
Real-world starters — SP core, DC fabric, interconnect, campus, secure edge, SD-WAN — each specified tightly enough to clear the clarify-gate and go straight to design.
The system keeps score on itself.
Across every run: claims grounded, hallucinations caught, the trust distribution, and which technologies the practice actually works — the honesty ledger, aggregated.
Compliance, honestly labeled.
A NIST CSF matrix mapping the design to every control area — and every row stamped “design-addressed — confirm by audit,” never pretending a design is an attestation.
The watch after the design.
A model-driven telemetry & SLO catalog, and a drift check of live read-only state against design intent. Alert-and-ticket only — the system never pushes to a device.
03Build 03 / 08
DWG P-03 · Developer tooling · research automation · live in production
AI Command Hub — a guide that researches itself.
A bilingual, Arabic-first reference for the Claude ecosystem: verified skills, vendor and community sources, and a library of 167 copy-ready commands. The catch: nobody maintains it by hand. A scheduled research agent scans GitHub trending and the official repos every day and files what it finds — trust-tiered, so you always know what is first-party and what to review before relying on it.
Scout 01
A daily scheduled agent sweeps fresh skills, subagents, MCP servers, and npm packages — and stamps the trusted ones, with the source of every finding visible.
Rank 02
Three explicit trust tiers — Verified first-party, official Vendor, and Community (“review before you rely on it”) — provenance as a first-class label, not a footnote.
Arm 03
Beyond the catalog: a roadmap builder that plans a project from an idea, ready advisor prompts, and 167 practical commands sorted into 13 working categories.
167Ready commands
37Skills catalogued
10Subagent patterns
AR/ENBilingual · daily refresh
The tour · 01 — the hub
One hub, the whole ecosystem, trust-tiered.
The front page states the deal: 18 official skills, 11 vendor, 8 community, 167 commands, 10 subagents — and three explicit trust tiers above the fold, so provenance is the first thing you read, not the last.
The tour · 02 — the daily scout
The feed that updates while you sleep.
A scheduled task scans the freshest skills, subagents, and MCP connectors every day at noon. Each finding arrives dated, versioned, source-attributed — and stamped trusted only when it comes from a verified publisher.
The tour · 03 — the toolkit
From idea to phased plan.
Type an idea in any language — or drop a package.json — and get a phased roadmap built only from catalog skills, with any missing capability flagged as a gap instead of papered over.
Prompts you can fire as-is.
Prepared, tested prompt cards — instant roadmap, senior-style code review, market scan, a 14-day launch plan — click one and the advisor opens with it loaded.
The verified shelf.
Eighteen first-party skills, grouped by what they are for — documents, design, development, enterprise — each entry carrying the verified badge that tier one demands.
167 commands, filed like tools.
The library opens on 28 programming commands — /architect to /accessibility — every one a copy-ready prompt with a one-line contract of what it does.
04Build 04 / 08
DWG P-04 · Enterprise ERP · multi-tenant intelligence · live in production
H-NERVE — an ERP with a brain of its own.
The studio's largest build: a white-label ERP intelligence platform, Arabic-first, running live for a real Jordanian group across hotels, dairy, agriculture, and education. Beneath every screen sits the Brain — a causal graph of the entire group, a council of specialist agents that debate each call, a what-if engine that simulates a decision before you commit to it, plus memory, benchmarks, and a narrator that greets the chairman each morning with what moved overnight. And it nests: every company inside the group runs its own full ERP within the group's. The Brain proposes; humans dispose — it is read-mostly by design, audited, and safe to let it self-tune weekly. It even keeps an IQ score on itself — recomputed every week from the measured accuracy of its own past calls.
Sense 01
A causal graph of the whole group — 659 nodes, 1,295 weighted edges. Every hotel, farm, and cohort a node; click one and the effect cascades through everything it touches, so the system reasons about consequences, not just totals.
Deliberate 02
Specialist agents argue each decision in council; the Brain weighs the dissent and drafts the recommendation with a confidence score — reasoning you can read, not a black-box number.
Simulate 03
A what-if engine turns the levers of the business — price, output, capex — and watches profit, risk, and Brain-IQ move in real time down the causal chain, or lets the Brain solve for the optimum itself.
659Causal nodes
1,295Weighted edges
127Brain-IQ · self-graded weekly
AR/ENArabic-first · full RTL
The tour · 01 — the group as a living system
The orrery: every domain in orbit.
H-NERVE opens as an orbit map — dashboard, brain, finance, team, and system circling the group's core, with a live clock and the Brain's current IQ in the corner. It greets you first with a morning brief: what moved while you slept, and the one signal that needs you today.
The tour · 02 — the brain, drawn
The causal graph — how the Brain actually sees the group.
659 nodes, 1,295 weighted edges: every entity in the group and every relationship between them. Click a node and the influence ripples outward through everything it touches. This is the substrate the whole platform reasons on — consequences, not just totals — carrying an 85% model-confidence score of its own.
The tour · 03 — the council
Pose one decision.
State the call in plain language — optionally focus it on a unit or an aspect (finance, supply chain, sustainability…) — then convene the council. It runs on an on-device reasoning engine, and every past session is archived with its confidence.
Hear every voice.
Five specialists — hospitality, agriculture, dairy, education, risk — each take a stance and defend it (supports · qualifies · opposes). The Brain sits at the centre, weighs the dissent, and drafts the recommendation. Reasoning you can read, not a black-box score.
The tour · 04 — simulate it, then look inside
Simulate the decision before you commit.
Drag any lever — room rate, dairy output, new capex, target return — and profit, revenue, risk, and Brain-IQ recompute live down the causal chain. Or let the Brain solve for the optimum, then hand the scenario straight to the council.
An ERP, inside the ERP.
Each company nests a complete operating console within the group's — intelligence, finance, operations, pipeline, and team in one place, opened here on the agriculture arm. A Company Health Index grades the company live across financial, operational, signal-load, and momentum axes, and every number rolls straight up into the group brain. Hotels, dairy, and education each carry the same.
The tour · 05 — the brain keeps score on itself
The Brain has an IQ — and it is climbing.
The platform grades its own intelligence. Every week the Brain reflects on the decisions it recommended, measures how they actually played out, and recomputes its IQ — 102 at launch, 127 today, at 77% measured accuracy. The self-report waits for a human to review and approve. An AI that keeps score on itself, in writing.
Five sectors, one nervous system.
The portfolio at a glance: hospitality, dairy, agriculture, education, and the holding itself — 2,065 employees and seven figures of 30-day revenue — each company a card, each card a door into its own nested ERP, all feeding the same brain.
The tour · 06 — the loop closes
A recommendation becomes a plan with steps.
The council's call does not end in a slide. The Brain drafts an executable plan — numbered steps, a confidence score, a live done-count — and a human commits it or abandons it. Steps tick off as the business works them, and the plan stays on the record either way: four plans here, ten steps, one already closed.
It forecasts what one company owes another.
The group's companies trade with each other — the hotels buy dairy from Al-Maha and produce from Loran. The Brain forecasts each flow a week ahead, quantity and all, shows the signal it reasoned from and how confident it is, and then waits: approve or reject. Nothing moves on the Brain's word alone.
The tour · 07 — nothing unproven, nothing forgotten
Every number is checked before you read it.
The contract is mechanical: every numeric claim the Brain writes must appear in the database payload that fed it, matched within a ±2% tolerance for editorial rounding. Anything that cannot be proven is flagged instead of published — and the trust badge on each line decomposes into four axes: verification coverage, data freshness, supporting density, causal-graph support.
The group's mistakes, kept on purpose.
Decision Memory is the institutional record — what was decided, what actually happened, and whether it worked — filed by sector, year, and outcome. A labneh batch written off after sitting visible on a dashboard for nine days. A group margin that compressed from 16.8% to 11.4% because nobody hedged energy. The Brain reads this before it advises.
The tour · 08 — measured against the world outside
It learns from peers it is never allowed to see.
Federated benchmarks across anonymized tenants: 16 connected peers, K-anonymity of 5 — no pattern surfaces unless at least five peers stand behind it, and any tenant can opt out at any time. The group gets to learn what six other mid-luxury MENA hotels did about heatwave discounting without a single one of them exposing a row.
Carbon and water on the same ledger as profit.
ESG is not a separate annual report here — it is scored quarterly per company, next to the financials: an average of 78 out of 100 across the group, with carbon, water, and renewable share tracked entity by entity, quarter by quarter. The same graph that reasons about margin reasons about the footprint.
05Build 05 / 08
DWG P-05 · LegalTech · grounded legal RAG · live in production
ISTINAD — legal AI that would rather refuse than guess.
استناد — “on the authority of the text.” Grounded legal research for Palestinian attorneys over 562 statutes of West Bank and Jordanian law, territory-tagged and amendment-aware. Every claim ships with citations and a confidence score; a deterministic grader — pure code, not a model — decides what may be cited, and when the material is not good enough, the tool refuses and says exactly what is missing. Decision-support for licensed lawyers, never a substitute for one.
Ground 01
The model only ever sees the graded, citable set of articles — and the submission gate mechanically rejects any citation outside it. No prose-only legal output path exists.
Grade 02
West Bank and Jordanian statutes read near-identically but differ in force. A deterministic grader checks territory tags and in-force status on every candidate — the famous “fake citation” failure mode, designed out.
Refuse 03
Below the confidence threshold the tool does not improvise — it declines, states what reliable material it lacks, and offers unverified discovery leads clearly marked as leads, not citations.
562Statutes indexed
14k+Articles · territory-tagged
2Legal territories · WB / JO
0Uncited claims possible
The tour · 01 — the research bench
Territory first. Question second.
The bench will not even take a question until the matter's territory is chosen — because West Bank and Jordanian statutes read near-identically while differing in force, and citing the wrong one is how lawyers get sanctioned. The disclaimer is the first thing on the page, on purpose.
The tour · 02 — the case & the council
The case, structured — never invented.
A real consumer-protection matter — rescission of an e-commerce sale before the Ramallah magistrate court — pasted in and typed out: parties, facts, claims, evidence, and relief, every item mechanically backed by a verbatim span from the file itself. Even the complexity meter is labeled honestly: input complexity only, never a verdict.
Three readings of one record.
Independent lenses — the judge's reading, the opponent's, the procedural one — deliberate the same graded record and surface where they disagree. Ungrounded positions are struck from the minutes, visibly.
The tour · 03 — بناء الاستراتيجية · the strategy, built the hard way
Four stages, none of them a shortcut.
Strategy is not one prompt. The bench derives the legal issues, retrieves statutes per issue, grades every retrieved article for reliability (CRAG), and only then synthesises — with the refusal rule printed on screen while it works: if the reliable material is not there, it declines. On purpose.
And when the record is thin — it says so.
The bench derived the five issues this file actually turns on — refund terms in e-commerce, the older-model delivery, the 40-day delay, the website's clauses against consumer rights — graded the citable articles behind each, and then did the rarest thing legal AI does: it abstained, named the exact issue it could not ground, and listed what material to bring. Refusal over invention — that is the entire product.
06Build 06 / 08
DWG P-06 · PropTech · conversational discovery · live demo
AQAR AI — a property search you talk to.
عقار AI — Arabic-first property discovery for the Amman market. Nobody looking for a flat thinks in dropdowns; they think in one sentence. So the search takes the sentence. A buyer writes «أبحث عن شقة للإيجار في عمّان بحدود ٦٠٠ دينار، غرفتين» and the assistant resolves it into a structured query — intent, type, ceiling price, bedroom count — runs it against the real listings, and then narrows by asking rather than by handing back a filter panel. Every property carries photographs and a 360° walkthrough, and browsing stays free for everyone; only reaching the seller asks for a name.
Parse 01
Ordinary Arabic — colloquial, with Eastern-Arabic numerals and mixed English — becomes typed search parameters. The parse is shown back to the buyer as chips they can see and correct, never applied invisibly.
Narrow 02
A partial answer is not a dead end. The assistant returns what already matches, then asks the one question that would cut the set down — district, price, size — and offers it as a tap, not a form.
Gate 03
Search, photographs and the full 360° tour need no account. The identity check sits on one edge only — contacting the seller — so the listing stays open while the lead stays real.
ARNative intent parsing
360°Walkthrough per listing
0Filters to fill in
1Gate · contact only
The tour · 01 — one sentence, resolved
The sentence is the query.
One line of ordinary Arabic — purpose, area, ceiling, bedrooms — comes back as visible chips above the thread: للإيجار · شقة · ≤ 600 · 2+ غرف. The buyer can read exactly what the machine understood before it answers, and the answer arrives with the next question already attached.
The tour · 02 — the listings, and the one gate
The panel is still there — for the people who want it.
Conversation is the front door, not the only door. The classic filter grid stays for a buyer who would rather set a price range by hand, and both routes resolve to the same query against the same listings.
Everything open, except the phone call.
Rooms, area, amenities and the full walkthrough are visible to anyone. The lock sits on a single edge — reaching the seller — and the page says so in plain words rather than trapping the tour behind a sign-up wall.
The tour · 03 — inside the property
A real sphere, on placeholder walls.
The walkthrough is a true equirectangular viewer — pan, tilt, zoom, and room-to-room hotspots (الصالة · غرفة النوم الرئيسية) that move the camera between spaces without leaving the page. The geometry in this build is generated placeholder interior, not photography: the pipeline is finished and waiting on real capture, and we would rather show it at that stage than stage a photograph we did not take.
07Build 07 / 08
DWG P-07 · EdTech · consented voice modelling · live demo
STUDY HALL — a teacher's voice, only with their permission.
A student stuck at eleven at night has nobody to ask until the next lesson. This gives them their own teacher's study assistant — one that explains the way that teacher explains, from the material that teacher actually uploaded. The hard part was never the model. It was building something that clones a real person without ever becoming a forgery of them: voice is measured arithmetically from the teacher's own writing rather than guessed at, cloning anyone else is refused without recorded consent, and the assistant states in its first breath that it is an AI — then refuses to write submittable homework, quote a mark, or say what is on the exam.
Consent 01
Cloning a colleague, a public figure or anyone who is not you requires an explicit recorded agreement. Without the tick the assistant is not created — the gate is mechanical, not a checkbox in a policy nobody reads.
Measure 02
Voice comes from things the teacher actually wrote — class messages, exam instructions, feedback. Rhythm, punctuation and Arabic/English mixing are measured from that corpus and scored against it afterwards. Arithmetic, not opinion.
Separate 03
How a teacher sounds and what a teacher knows are two different inputs, kept apart by design. A student's question pulls in the matching unit of course material; it never changes the voice, and the voice never invents the material.
0Clones without consent
2Inputs · voice ≠ content
AR/ENAnswers in the asked language
SelfVoice distance scored
The tour · 01 — the first thing a student reads
The disclosure is the headline, not the footnote.
Before the class code, before the teacher list: “This is an AI, not your teacher.” It will not write the homework, and it cannot tell a student their marks, their deadlines, or what is on the exam — ask the teacher for anything that counts. Written where a fourteen-year-old will actually read it.
The tour · 02 — consent, then voice, then content
Choose “another teacher”, and a gate appears.
The moment the subject of the clone is somebody other than you, the form grows a requirement: they have agreed to this clone. The console says outright that this is not a formality — untick it and nothing is built. Consent as a control in the code, not a line in the terms.
How you sound. What you know. Two boxes, never one.
Section II takes things the teacher wrote — “i said this three times in class — write the units. no units, no mark.” Section III takes what they teach, one named unit per block. The question pulls the unit; the corpus sets the voice. Neither is allowed to contaminate the other, which is precisely why the result sounds like a person instead of a chatbot wearing their name.
08Build 08 / 08 · sealed
DWG P-08 · Archforge Internal · access sealed
ARCHGATE — the build behind the builds.
Every system above was drawn faster and truer than it had any right to be. This is why — and it is the one build we will not lay on the table. Archgate is the studio's own instrument — an agent harness, the same discipline of machinery that runs the studio's 130+ task-specialised subagents: a place where the moving parts of an intelligent system stop being code you rewrite and become assets you can version, weigh against each other, and re-forge on command. We won't name its parts, open its other rooms, or explain how the pieces lock. What we will say is small and true — it composes, it remembers, it grades its own work, and it runs entirely on the edge. Everything else stays in the vault. A workshop that means it keeps one tool off the table.
Composed 01
Nothing here is written once and frozen. Every part is an asset with a version and a history, recombined into new systems without starting over. That is the trick — and as far as the trick gets described.
Weighed 02
Nothing leaves this room unmeasured. The instrument scores its own output against a standard before any of it is trusted — so “it works” is a number, not an opinion.
Sealed 03
You are seeing one screen and no more. The other rooms, the parts, the wiring — withheld by design. What vouches for the machine is the machines it has already built, above.
▮▮Assets under version
▮▮▮Live compositions
SelfEvery run graded
EdgeGlobal · zero cold starts
The tour · 01 — one glimpse, and no more
The overview — and nothing behind it.
One screen: the shape of the instrument at altitude, its lattice lit, everything around it blurred on purpose. The other sections are sealed, the values redacted, the labels withheld. You are meant to see that it runs — not how. That is the whole reason it is shown this way.
That closes the ledger — eight builds, one kept behind glass. The instrument that shapes them does not go on the page; if that is what you came for, the conversation starts privately.