Geneva / Full-stack development / Remote-friendly

Dylan Buchi

I turn messy workflows into products people can trust.

I am an early-career full-stack developer who designs and ships focused SaaS products end to end, from interface and authentication to data, integrations, deployment, and verification.

Explore Casebound Start a conversation

Seeking an early-career full-stack engineering role

Departures

Recent work on the board

Track	Destination	Status
01	Casebound	PayPal AI submission
02	CloseChase	H0 submission
03	INVISIBLE // 2126	Category winner
04	Open Here	Hackathon project
05	HSBG Mac Overlay	Open source

Product interfaces
End-to-end systems
Safe integrations
Verified releases

Scroll to the work

01 / Lead case studies

One painful workflow. One complete loop.

Casebound and CloseChase show how I work: define the workflow, model its states and permissions, ship the complete loop, and verify what is actually ready.

Casebound

PayPal AI Hackathon submissionAI agent product engineering

An AI dispute desk for PayPal sellers. A Claude agent investigates each dispute, works out whether fighting is worth the money, and drafts the response. Nothing is sent until the seller approves.

Hackathon submissionPayPal AI Hackathon 2026 (Devpost)

Open the live demo View source Watch the demo

Casebound verdict for a $189 item-not-received dispute: the agent recommends a refund, compares refunding with fighting, and waits for the seller to approve the drafted reply

Agent trace / sandbox case

The agent is working the case

Reading the dispute on PayPal
Opening the shop’s fulfillment record
Working lean Fight
The shop record shows a signed delivery.
Checking PayPal shipment tracking
Changed my mind Accept
PayPal tracking says RETURNED, which contradicts the shop record.
Running the money math

The starting point

A PayPal dispute means checking the dispute, the payment, shipment tracking, shop records, store policy, and buyer history, then deciding whether a fight is worth it. Casebound runs that investigation and shows its reasoning, so the seller decides with the evidence in front of them.

The route through a dispute

01Investigate the case
The agent reads the dispute, payment, tracking, shop records, store policy, and buyer history through tools.
02Weigh the money
Expected-value math compares refunding with fighting. Tiny disputes get a quick call from plain math, with no AI.
03Change course on evidence
When sources disagree, the agent can change its mind mid-investigation and explain why.
04Draft, then approve
It drafts the response. Nothing goes to PayPal until the seller approves it.

How it is built

Interface: Next.js 16 / TypeScript / AG Grid
Agent: Anthropic Claude with tool use
PayPal: Disputes API / Orders & Captures / shipment tracking
Checkout: PayPal JS SDK (sandbox)
Hosting: Render / Cloudflare DNS

How it is checked

73 automated tests
AI evals: synthetic cases 20/21, held-out messy cases 12/15, prompt injection 2/2
Code-enforced source grounding, plus a draft self-check that rewrites unsupported claims
Security review with fixes: rate limits, signed passcode cookie, server-side checks
Multi-agent code review, with fixes applied

Current boundary

PayPal sandbox only: no real seller accounts or real money
Hackathon submission with no result yet; judging starts after Nov 12, 2026
A new build: first commit Oct 28, 2026, live Oct 29

CloseChase

H0 hackathon submissionFull-stack product engineering

A deployed product that helps bookkeeping firms collect explanations and receipt context for unclear transactions before month-end close.

Hackathon submissionH0: Hack the Zero Stack with Vercel v0 and AWS Databases

Open live product Read the Devpost submission Watch the demo

CloseChase public homepage showing an illustrative month-end transaction request batch

Illustrative workflow

Alpine Café GmbHCurrent close / due in 7 days

Answered: 7/12
Receipts: 3
Open: 5

UberCHF 42.80Answered
GalaxusCHF 189.90Receipt missing
Booking.comCHF 612.30Needs explanation
SwisscomCHF 129.00Ready

The starting point

Unclear transactions, missing receipts, and vague client replies are scattered across inboxes. CloseChase turns that chase into one tracked request batch.

The route through the product

01Create request batch
Import or add unclear transactions for one close period.
02Share private link
Generate a private client portal link, shown once, and send it through the firm’s existing channel.
03Collect context
Clients answer without an account and add demo-safe receipt context.
04Review and close
Approve or follow up, export close notes, then close or revoke the portal.

How it is built

Interface: Next.js App Router / Vercel
Operator access: Clerk Organizations
Workflow state: Amazon DynamoDB
Receipt binaries: Private S3-compatible storage
Verification: Playwright / fake-data smoke checks

What works now

Firm-scoped operator workspace and client directory
Batch, transaction, portal, review, activity, and export states
Passwordless private client portal using token-hash lookup
CSV import/export with formula-injection mitigation
Demo-safe receipt upload/download with split metadata and binaries

Current boundary

Controlled fake-data pilot workflow, not ready for real customer receipt data
No active billing or accounting-system write-back
No AI/OCR or real email reminder delivery
Real receipt readiness still needs malware scanning, retention/deletion policy, audit logging, and legal/privacy review

02 / Selected proof

Selected work

Public work shows range; private work is labeled clearly.

01
INVISIBLE // 2126
Award-winning web experience
A3 Web Design Challenge 2126 Portfolio category winner / 4th overall / 198 of 300
A classified archive from 2126 with a redacted identity, coded projects, and an interactive terminal, built under a 48-hour design challenge.
- Next.js
- Tailwind CSS
- Interaction design
- Responsive animation
Open project View source
02
Open Here
Qloo Agentic Hackathon project
An AI agent that reads Qloo taste data to map where a business idea has hidden demand, then plans the launch around that crowd. It scores its picks against a plain-AI baseline with the same measure and checks every artist and brand against what Qloo returned.
- Node.js
- LLM tool calling
- Qloo API
- Leaflet
- Vercel
Open project View source
03
HSBG Mac Overlay
Open-source macOS app
A native overlay for Hearthstone Battlegrounds that labels each leaderboard player with rating, tier, combat record, streak, and tribes, parsed live from the game log. MIT licensed, released as a universal build from GitHub Actions.
- Swift
- SwiftUI
- AppKit
- Real-time log parsing
Open project View source
04
Codex Workflow Kit
OpenAI Build Week submission
A private-source, repo-native control and evidence layer for serious Codex development: scoped tasks, bounded workers, independent review evidence, and closeout gates that reject stale proof.
- Node.js
- CLI tooling
- Git hooks
- Test gates
Open project Watch the demoPrivate repository / public architecture summary
05
Travelio
Full-stack booking experience
A marketplace-style stays app with profiles, property management, favorites, reservations, and trip tracking.
- Next.js
- TypeScript
- MongoDB
- Prisma
Open project View source
06
Big Movies
Front-end product surface
A movie discovery app with auth, search, watchlists, favorites, trailers, and actor profiles using the TMDB API.
- React
- Redux
- TypeScript
- Material UI
- TMDB API
Open project View source
07
Folder Synchronizer
Python automation utility
A predictable one-way source-to-replica folder sync tool with command-line operation and explicit logging.
- Python
- File I/O
- CLI
- Automation
View source

03 / Experience

Experience

Product work, team delivery, open source, and freelance automation.

Sep 2024 - Present
Geneva / Remote
Independent Software Developer
Personal Projects
- Independently designed and built personal SaaS workflows across interface, authentication, data, integrations, deployment, and verification.
- Built and shipped Casebound for the PayPal AI Hackathon 2026: a Claude tool-use agent that investigates PayPal disputes through the Disputes, Orders, and tracking APIs and drafts replies the seller approves; 73 tests, AI evals, and a security review; live in the PayPal sandbox.
- Built CloseChase as a deployed H0 hackathon submission using Next.js, Clerk Organizations, DynamoDB, private object storage, CSV flows, Playwright, and Vercel; controlled fake-data pilot.
- Built Open Here, an AI agent on Qloo taste data for the Qloo Agentic Hackathon, and released HSBG Mac Overlay, an open-source native macOS app with a GitHub Actions universal build.
Jun 2024 - Aug 2024
Lausanne / Hybrid
Software Engineer Intern
acheteur.ch
- Contributed React and Next.js components to a shared TypeScript monorepo for a large-scale real estate product.
- Translated Figma work into reusable UI and delivered frontend features and refactors in an Agile team.
Jul 2019 - Jun 2024
Geneva
Self-Taught Software Engineer
Independent study and projects
- Built a broad engineering foundation through public projects spanning web applications, Python automation, databases, data structures, and system design.
- Progressed from small utilities to deployed full-stack applications with authentication, persistent data, and external integrations.
2020 - 2021
Remote
Open Source and Freelance Developer
The Algorithms / SwissBorg
- Merged 13 Python pull requests into The Algorithms, improving typing and resolving mypy dependency issues.
- Built a SwissBorg Discord bot with automated responses, admin controls, uptime checks, and live information retrieval.

04 / Capabilities

Capabilities

The stack changes; the job is to connect the interface, data, permissions, integrations, and release around one useful workflow.

Product interface

Turn workflow states into clear, responsive surfaces with deliberate information hierarchy.

React
Next.js
TypeScript
Tailwind CSS
Figma implementation
SwiftUI

Full-stack systems

Model the data, permissions, and server behavior needed to carry the workflow end to end.

Node.js
Python
Swift
DynamoDB
PostgreSQL
SQLite
REST APIs

Safe integrations

Keep organization access, private links, storage, and third-party services scoped and predictable.

Clerk
OAuth
PayPal APIs
Claude tool use
Private object storage
CSV safety
Provider validation

Verification and release

Test the real workflow, failure paths, and deployed surface before calling it ready.

Playwright
Vitest
AI evals
Vercel
Render
Smoke checks
Health checks
Git

05 / Principles

How I work

Three constraints keep the work focused and honest.

Focus the workflow
Solve one painful process completely before expanding the platform.
Make limits visible
Permissions, private links, storage, and third-party limits should be clear in both behavior and copy.
Verify what ships
Tests, browser checks, and honest product boundaries matter as much as the deployment.

06 / Contact

Have a messy workflow worth fixing?

Tell me what is stuck, who it affects, and what a useful outcome looks like.

Emaildev@dylanbuchi.com

I turn messy workflows into products people can trust.

Departures

Casebound

The starting point

The route through a dispute

How it is built

How it is checked

Current boundary

CloseChase

The starting point

The route through the product

How it is built

What works now

Current boundary

Independent Software Developer

Software Engineer Intern

Self-Taught Software Engineer

Open Source and Freelance Developer

Product interface

Full-stack systems

Safe integrations

Verification and release

Focus the workflow

Make limits visible

Verify what ships