Automation for the part of admissions that is pure noticing
The follow-upthat happens anyway
The reference that has not come back. The registration form filled in twice, under two spellings of one surname. All of it has to be noticed on the right day, by somebody, every time. A program can do the noticing, draft the reminder and match the duplicate. A person still reads it before it goes. Nothing about a child gets decided by a model.
Book a 15-minute callSend the details below. I'll reply by email to arrange a time.
London, UK · Available for new projects
What actually slows independent schools
The registrar's inbox is half the admissions system
iSAMS or Engage has the registration and the fee. The open morning list came off a web form, the assessment results are on a spreadsheet, and the chasing sits in the registrar's sent items. Every one of those is a separate family record with a different spelling of the same name.
The number the governors ask for lives in Excel
The standard reports in SIMS or iSAMS answer the questions somebody anticipated. The one the finance committee asked for, cut the way they asked for it, becomes an export to a spreadsheet and a set of formulas nobody else can check. So the same export gets rebuilt before every meeting.
One trip, three systems, four messages
A trip or a club sign-up is a letter by email, a consent form on Firefly, and a payment on ParentPay. The class WhatsApp group carries a fourth version that nobody at the school wrote. Whoever missed one of them rings the office.
Assessing a bursary means opening whatever the family sent
Payslips as photographs, accounts as PDFs, a form somebody filled in by hand and scanned. It lands in an inbox, gets forwarded to the people assessing it, and a decision has to be recorded and defended later. Family financial documents sitting in a shared drive is nobody's intention, and under UK GDPR it matters that you can say who opened them.
AI and automation
Retrieval over your own documents is where the work happens, so the answers come out of your material. The model is rarely the useful part. Evals catch a regression before a customer does, and jobs shout when they fail. Where a scheduled script would do the job better, I will say so and write the script.
- A list of what not to automate
- The first pass is an audit, and some of what comes back is a recommendation to leave it alone or replace it with a scheduled script. You get that list either way, including the parts that would have been billable.
- Answers that come from your own material
- Retrieval over your documents, your policies and your history. When it cannot find an answer in your material it says so.
- Alarms that mean something
- Every job reports whether it ran, and a failure wakes somebody up. A process that dies quietly while the dashboard stays green is worse than one that never started.
Selected projects
Products I built and open-source projects I maintain. Open a project to see the work for yourself.
- 012026
Dr Flavia Pretti Aslanian
Healthcare · Practice site
A bilingual dermatology website for a practice in Harley Street and Wimbledon
A new landing page guides patients to medical, aesthetic and surgical care. The practice edits its own content, and staff have a password-protected dictation tool in the same app.
- Next.js
- Sanity CMS
- Tailwind v4
- Radix
- Motion
- ISR
- Railway
- Cloudflare
Design · Build · Deploy
- 84 published English and Portuguese routes
- 72 treatment pages with clinic-editable content
- Custom before/after comparison UI
- Internal clinician dictation tool
- 02
2025 →Delta Xi
E-commerce · Brand
A mathematics apparel brand on a storefront I wrote from scratch
I built the cart, checkout, product recommendations, affiliate programme and support desk, and run the infrastructure myself.
- Next.js
- React Server Components
- Stripe
- Auth.js
- Tailwind
- Umami
- Chatwoot
- Cloudflare
Founder · Engineer · Operator
- Custom checkout on Stripe + Klarna
- Server-rendered recommendation engine
- Self-hosted analytics and live chat
- 50+ STEM explainers driving organic search
See 4 more projects
- 03
2026Continuum
Web3 · DeFi protocol
A Solana DeFi protocol for 24/7 synthetic exposure to real-world assets
Deposit USDC, mint matched long and short tokens at NAV, sell the leg you don't want. A constant-product invariant keeps the pair redeemable back to the deposit.
- Rust
- Anchor
- Solana
- TypeScript
- Pyth
- Next.js
- Bun
- Railway
Pre-mainnet. Devnet markets, audits pending.
Protocol · Frontend · Infra
- 5 Solana smart contracts in Rust
- 8 synthetic markets on devnet
- 4-state oracle risk machine
- Volatility-scaled fee curve
- 042026
Dray
AI tools · Product website
A product website for keeping project context across AI agents
An editorial landing page introduces shared project context, shows illustrative agent workflows and explains setup and privacy ahead of beta.
Beta coming soon. The public demos illustrate the workflow; they are not live agent connections.
Design · Build
- Interactive agent-linking illustration
- Project-context walkthrough
- Setup and privacy guidance
- 05
2026Tallylamp
Open source · Browser infrastructure
A browser for AI agents that you can watch and take over
I built the browser service, dashboard, MCP server and public website. Agents use Chrome on your server while you watch the same browser live. Take control to sign in, then hand it back; saved profiles carry between sessions.
- TypeScript
- Next.js
- Chrome
- MCP
- SQLite
- Docker
- Railway
Public site at tallylamp.dev. The browser service and dashboard are self-hosted.
Creator · Engineer · Maintainer
- Live viewing and explicit human handoff
- Separate agent permissions and reusable profiles
- MIT licensed, with a Railway deploy template
- 06
2025 →CarClout
SaaS · AI content engine
A subscription app that turns a phone photo of a car into something worth posting
Upload a photo, pick a template, get the edit back a moment later. Thirteen AI models sit behind a metered credit ledger, next to a layer editor, a post scheduler and a live chat room.
- Next.js
- SurrealDB
- Auth.js
- Stripe
- fal.ai
- Cloudflare R2
- Stream.io
- Railway
Engineering · Infra · SEO
- 724 generated landing pages, all live
- 128 API routes behind one app
- 13 AI models on a metered credit ledger
- Instagram posting, scheduling and analytics
One runs on the client's own domain and one is internal to theirs, so there is nothing here of mine to show you.
- ClientJooyontechAI sales assistant, Korean and English
- ClientPHFWeb, AI and automation · private medical practices, UK
Not everything is public. A fair amount of the work sits in private repositories I cannot hand you a link to. Ask, and I will walk you through it.
Ask to see moreHow a project runs
- 01Before code
Scope
One call, then a written spec. What it does, what it deliberately does not do, and what milestone one is. You sign that off before I write a line.
- 02Early
First deploy
A narrow slice that actually works, running where it belongs. A URL you can open, a build on your phone, a job on your own servers. Real data, real hosting, no placeholder screens.
- 03Ongoing
Build in the open
Deploys land as the thing fills in, not in one drop at the end. You watch it happen. And you get to change your mind while changing it is still cheap.
- 04Handover
Yours to run
The product runs in your accounts, with the domain, hosting and documentation under your control. Source-code transfer is quoted separately; we agree what is included during scoping.
Rafael Aslanian
I got my hands on a computer at 6 and never really stopped.
Now it is decentralised finance, tokenising real-world assets and topological deep learning, then building whatever those turn into. Continuum came out of that. So did Delta Xi (δξ, an infinitesimal change in a random variable).
Bossa nova and MPB were always on at home, so I sang before I played anything. Guitar and piano came later, then jazz.
- Based
- London, UK
- Languages
- English, Portuguese; intermediate Spanish and Chinese
- Focus
- Web3, DeFi, real-world asset tokenisation
- Also
- Music
Have an idea? Tell me.
Need the audience too?
I can build the thing. Getting people to notice it is a separate job, and not mine. For that I work with Nytforge, who film, edit and put the work in front of a local audience. They are strongest with automotive and local businesses.