KHOURY/ atelier
khoury.uk · a boutique AI studio

Bespoke AI,
built with a catch.

We build bespoke AI for the work that actually matters — and we never ship it without a catch. Every tool pairs genuinely powerful intelligence with a deliberate guardrail it can't talk its way past, so you get all of the upside and keep complete control.

Explore the tools ↓ See why they're safe
In the workshop — Watchdog FPL Predictor Steward Job Assistant Headless Orchestrator Aider Lab Provost
The atelier

Anyone can give you raw power.
We give you control.

Generic AI is impressive — and often unreliable or unsafe for real work. The hard part, the part worth paying a specialist for, is knowing exactly where to put the brake, the gate, and the second pair of eyes. Every Khoury build is genuinely capable, yet refuses to be talked out of its limits.

01 — privacy by design

Your data stays home

Your tools run on your own hardware for the everyday work. Nothing leaves your house without your explicit, case-by-case permission.

02 — the catch

A guardrail per build

Every product carries one signature safeguard, tuned to its job: a human approval, a rule it cannot break, an undo button. The catch is the feature.

03 — hand-tuned

Built for your reality

Every tool is hand-built, made well, and run on real hardware — not a hundred off-the-shelf wrappers. Each is tuned to the machine it lives on and the person it serves.

The collection

Bespoke systems. Built with a catch.

From private financial confidants to autonomous fleet telemetry and quantitative prediction engines. Each card names the safety catch it ships with.

★ FLAGSHIP
No. 01 — operations · telemetry

Watchdog

the continuous safety radar for running AI fleets

A real-time cockpit and invariant sentinel running on your own server. It tracks host vitals, containerized agent vessels, task pipelines and in-flight workspace changes at a glance — ensuring multiple autonomous systems never run blind.

⛓THE CATCH — zero cloud telemetry. Built on Python standard library with zero third-party packages, it trips an immediate hard stop the moment a credential leak, swap spike, or rogue change is detected.
Open the dossier → quantitative sports
No. 02

FPL Predictor

closed-form expected points & 3-GW horizon transfer solver

A 91-task quantitative forecasting engine for Fantasy Premier League. Decomposes Poisson goal and clean-sheet densities into minutes-conditioned expected points, discounted at γ = 0.90 across a 3-GW rolling window with exact branch-and-bound integer programming.

⛓THE CATCH — zero invented priors. Projections are mathematically grounded in verified historical snapshots and market densities — no subjective model drift or vibes.
Open the dossier → finance
No. 03

Steward

the private financial confidant

A private companion for the numbers you'd never paste into a normal chatbot. It runs on your own machine, doing the day-to-day quietly; the hard questions only ever go further with your explicit say-so.

⛓THE CATCH — your financial data never leaves your house without your say-so — and it never runs unattended.
Open the dossier → white-label
No. 04

Job Assistant

a job-search service under your own brand

Stop hunting. Start choosing. The right roles, matched to you and drafted from your real experience — ready the moment you wake.

⛓THE CATCH — it can't invent a word of your experience, and it never applies for you. The decisions — and the send — stay yours.
Open → white-label
No. 05

Headless Orchestrator

your brand's intelligent engine

An engine that runs whole teams of AI agents under your own brand, handling complex workflows reliably — without surprise bills or rogue moves.

⛓THE CATCH — it follows fixed rules, not vibes — with hard spending and permission limits, it stops rather than guesses.
Open → software
No. 06

Aider Lab

a private helper that stays home

A patient helper that builds and fixes software for you — and does all of it on your own computer, so your work stays at home. When it wants a second opinion, it can ask a cleverer expert outside — but only if you turn the key.

⛓THE CATCH — kept private by default. The outside expert is only ever called in when you unlock the door — and even then, your work itself never leaves the house. It's yours, start to finish.
Open the dossier → in design
No. 07

Provost

the senior partner that reasons, then builds

Bring a raw, half-formed idea. Provost interrogates your thinking the way a senior partner would — three to five rounds of bias-aware questioning that pressure-test the assumptions behind expensive mistakes — then routes the work through a hierarchy of senior agents that plan, build and QA it, every handoff checked and gated.

⛓THE CATCH — no model can wave itself through. Every deciding gate is deterministic code, and the first gate is your approval.
Open the dossier →
Under the hood

Where the AI ends
and the catch begins.

Input flows left to right through the AI — then hits the catch: the emerald gate that decides whether anything proceeds.

input the AI doing the work ⛓ the catch — a human or a hard rule output
No. 01 · operations · telemetry
the continuous safety radar for running AI fleets

Watchdog

When you run multiple autonomous agents, you cannot afford blind spots. Watchdog is an autonomous fleet cockpit that watches your host vitals, containerized agent vessels, DAG task graphs, and workspace hygiene in real time without sending a single byte of telemetry to third parties.

⛓ THE CATCH100% local-first telemetry. Built on standard Python with zero external packages, it trips deterministic safety tripwires if unquarantined credentials, memory leaks, or boot limit violations are detected.
watchdog.py — nixbox atelier cockpit (v25.11) 
  ⬡ NIXBOX ATELIER COCKPIT — Fleet Radar & Invariant Radar
  ──────────────────────────────────────────────────────────
  Host: nixos (25.11) │ Tailnet: 100.111.179.49 │ CPUs: 4
  RAM:  [█░░░░░░░░░░░░░] 3.2/30.3 GB (zram: ON) │ Disk: 375GB free
  Fleet: ● job-assistant · ● dev-sandbox · ● orbridge :8765
  ──────────────────────────────────────────────────────────
  💼 CAREER INTELLIGENCE & AUTOMATION
    Job Search Assistant  ─  ✓ clean [docker:job-assistant ●]
  🎯 QUANTITATIVE PREDICTION ENGINES
    FPL Predictor git:main  ─  ▲ in-flight (⚡ 91 tasks: 54 done)
    World Cup MCMC Model   ─  ✓ clean (teletext render engine)
  🤖 AUTONOMOUS AGENT VESSELS & HARNESSES
    STEWARD Agent Services ─  ✓ clean [docker:dev-sandbox ●]
    AIDER Coding Lab       ─  ✓ clean [docker:ubuntu-dev ○]
    Headless Orchestrator  ─  ✓ clean (stateless protocol)
  ⚡ AI GATEWAYS & LIVE WEB ENDPOINTS
    ORBRIDGE Gateway :8765 ─  ✓ clean (token firewall)
    Khoury.uk Production   ─  ✓ clean (Cloudflare edge)
  ──────────────────────────────────────────────────────────
  🛡️  GUARDRAIL SENTINEL RADAR
    [✓ PASS] Secret Firewall (.env Quarantine) → clean
    [✓ PASS] Boot Generation Limit (Cap <= 3)  → 3 gens
    [✓ PASS] zram Swap Safety Cushion          → Active (7.5 GB)
  ──────────────────────────────────────────────────────────
  Fleet Pulse: ● clean · Commands: watchdog -w · watchdog -d
No. 02 · quantitative sports · optimisation
closed-form expected points & 3-GW horizon transfer solver

FPL Predictor

A 91-task quantitative forecasting engine for Fantasy Premier League. It eliminates emotional bias and recency hallucinations by decomposing player performance into closed-form Poisson and negative binomial scoring rates, conditioned on playing-time exposure probabilities. A multi-period horizon beam search (3-GW discount decay γ = 0.90) convolving 8 legal formations solves optimal starting XIs and transfer paths under exact branch-and-bound legality certificates.

⛓ THE CATCHZero invented priors. Every projection is grounded in immutable raw match snapshots and published market densities. The model refuses subjective overrides, enforcing exact mathematical proofs before any squad recommendation is issued.
Input
Raw Snapshots
immutable Opta lake
→
Engine
Poisson xP
closed-form terms
→
Minutes
Exposure Gate
bivariate π1/π2
→
Solver
3-GW Horizon DP
decay γ = 0.90
→
Proof
Legality Gate
branch & bound D47
→
Output
Ceefax Squad
weekly deliverable
P302   CEEFAX 302   Sat 15 Aug   19:30/21 KHOURY TELETEXT
GW1 OPTIMAL SQUAD · 3-GW HORIZON (DECAY γ=0.90)
FORMATION: 3-5-2 SQUAD COST: £98.5M (£1.5M ITB) GW1 xP: 66.94 3-GW EXP: 177.28 xP
KICKOFF PICKS: SAKA [C] · HAALAND [VC] · MBEUMO
STARTING XI (3-5-2) 58.15 xP LINEUP
SUN Roefs
GKP · £5.0m 3.56 xP
ARS Calafiori
DEF · £5.5m 4.86 xP
IPS O'Shea
DEF · £4.0m 4.58 xP
IPS Furlong
DEF · £4.0m 4.57 xP
ARS Saka [C]
MID · £9.5m 13.38 xP
MUN Mbeumo
MID · £8.0m 5.83 xP
ARS Eze
MID · £6.5m 5.80 xP
MUN Cunha
MID · £8.0m 5.65 xP
MCI Foden
MID · £7.0m 5.47 xP
MCI Haaland [VC]
FWD · £15.5m 7.19 xP
TOT Richarlison
FWD · £6.0m 4.58 xP
ORDERED BENCH RESTRICTION D14
TOT Dubravka GK
£4.0m 3.34 xP
SUN Ballard S1
£5.0m 4.53 xP
MUN Maguire S2
£5.0m 4.39 xP
BRE Wilson S3
£5.5m 4.39 xP
3-GW HORIZON MATRIX γ = 0.90
• Beam width: 20 paths evaluated
• Amortized transfer cost: -4.0 xP/hit
• Bench auto-sub weight: w = [0.21, 0.06, 0.002]
• Legality certificate: D47 Branch & Bound PASS
CLUBS (15): ARS 3 MUN 3 IPS 2 SUN 2 MCI 2 TOT 2 BRE 1
100 SQUAD
200 3-GW TRANSFERS
300 PROJECTIONS (xP)
400 DEADLINE 2026-08-21
★ UPDATED WEEKLY BEFORE EACH GAMEWEEK DEADLINE
No. 03 · finance · private
the private financial confidant

Steward

Steward lives on your own computer, not in the cloud — handling everyday bookkeeping and analysis in private. If a question needs a bigger AI brain, it strips out your personal details first and only sends the anonymous version once you say yes — twice. You get elite financial counsel with none of the exposure.

⛓ THE CATCHYour financial data never leaves your house without two explicit approvals — and your identity never leaves at all. It never runs unattended.
Input
Your numbers
on your machine
→
AI · private
Works privately
daily analysis
→
You
You approve
for hard questions
→
Privacy
Identity hidden
data anonymised
→
AI · expert
Trusted answer
elite counsel
→
Output
Clarity
back home
No. 04 · careers · white-label
a job-search service under your own brand

Job Assistant

Job Assistant lifts the invisible grind of a job hunt — honestly. It works the whole market for you: surfacing roles you'd never have found, weighing each against your real experience, and drafting honest, tailored applications for the ones worth your time. Each morning the shortlist is ready — you just choose, and send.

⛓ THE CATCHIt can't invent a word of your experience, and it never applies for you. It does the legwork; the decisions — and the send — stay yours.
Input
Your goals
drawn out
→
AI
Tailored search
matched to you
→
AI
Found + ranked
why each fits
→
Guard
Real experience only
no fabrication
→
You
You choose & send
always your call
→
Output
Shortlist
every morning
No. 05 · white-label
your brand's intelligent engine

Headless Orchestrator

A request comes in, and a fixed plan — set in advance, not improvised — coordinates a whole team of AI agents to do the work and bring the results back. Spending, scale and permissions are hard ceilings, and anything that needs your sign-off simply stops and waits.

⛓ THE CATCHIt follows strict rules, not vibes. The moment it reaches a spending limit or a permission boundary, it stops and asks — it never guesses.
Input
Your request
client task
→
Plan
Fixed plan made
set steps
→
AI · team
Agents do work
in parallel
→
Limits
Safety check
budget + permission
→
Output
Final results
under your badge
No. 06 · software · dev tooling
a private helper that stays home

Aider Lab

Aider Lab is a quiet, patient helper that builds and fixes software for you — running entirely on your own computer. Nothing is sent out; your work simply stays home. You tell it what you'd like in plain words, and it gets on with the job in its own unhurried way. If it ever wants a wiser second opinion, it can call on a cleverer expert outside — but only after you unlock the door, and even then through one watched channel. And before it keeps any change, it shows you exactly what it did — so you can say yes, or undo it in a heartbeat.

⛓ THE CATCHEverything stays home by default. The outside expert is only invited in when you open the door yourself — and nothing is kept until you've said yes. The final say is always yours.
Input
Just ask
in plain words
→
Helper
It does the work
on your computer
→
Home
Stays private
you turn the key
→
Expert
A wiser check
only if you allow
→
You
You say yes
or undo it
→
Done
All finished
kept at home
No. 07 · strategy · in design
the senior partner that reasons, then builds

Provost

Provost is a multi-agent software pipeline, currently in design — an entire software firm rendered as one line. It opens like a sharp consultant: a Socratic interview that challenges you three to five times and names the cognitive biases — anchoring, sunk cost, the planning fallacy — that quietly sink projects. Nothing is built until you approve the plan in plain English. Then the hierarchy runs: a Senior Planner decomposes the work and fans it out to specialists, a Senior Builder refuses copy-pasted or unreviewed code, a Senior QA probes spec drift, edge cases and missing tests — while a Bandit Watchdog patrols every run for drift, stalled agents and runaway cost, functioning as a strict health monitor rather than an optimization engine.

⛓ THE CATCHHandoffs are checked against explicit, deterministic criteria, and the gates that decide are LLM-free code — never an AI grading another AI. The line halts until a real approval is written, and the very first approval is always yours.
Input
Your raw idea
half-formed
→
AI
Socratic interview
3–5 bias rounds
→
You
Approve the plan
plain English
→
AI · firm
Senior agents
plan · build · QA
→
Code
Deterministic gate
LLM-free
→
Output
Reviewed software
shipped
The house rule

We don't sell raw power.
We sell absolute control.

Raw capability is a liability if you can't trust it. That's why we never build anything without a deliberate safety catch wired in — the difference between an impressive science project and a dependable tool you can confidently hand to your team. Tell us the problem; we'll design the catch.

Commission a build → Browse the collection