arrow_back Back to portfolio
smart_toy AI Lab · ongoing

AI Applications & Initiatives

Building practical AI systems from 0 to 1. This page features live experiments and structured workflows, spanning autonomous agents and robust frameworks that solve real-world needs and boost productivity.

Live demo · hackathon build
September 2026

Public demo with challenge tickets · Open Live Demo — start on the desk (try #8)

Service Desk Copilot

Operational service desks waste time on mislabeled tickets — this sidecar drafts the route and reply, then waits for a human.

Python Rules engine Chroma OpenAI Apertus Agent tools RAG Vercel

Service desk analysts at Swiss Life operational desks still decide every ticket. Service Desk Copilot is a sidecar for that queue: it reads an incoming Jira-style ticket, proposes seven triage fields, pulls similar resolved cases, and drafts a requester comment. Rules and a service catalog resolve service, team, and priority before any chat model runs; retrieval corroborates; a light LLM fills low-confidence fields and a heavier model writes the draft behind a code gate. The analyst approves, edits, or rejects—nothing auto-closes. Built solo in 36 hours for the Swiss Life challenge at Swiss {ai} Weeks Zurich Hackathon 2026 (Sep 24–25), with an OpenAI default path and Apertus as a Swiss open-weights option. On the challenge set it surfaces 4/20 mislabel traps, cuts review reading volume ~28% vs full ticket text, and rules-routes in under a second.

rule

Rules First

Catalog and body tokens route service, team, and priority before any chat call.

draft

Evidence Draft

Similar cases and a gated requester comment sit next to the decision card.

verified_user

Human Loop

Approve, edit, or reject — feedback writes back; tickets never auto-close.

Live demo · iterating
August 2026

Public demo with sample data · Open Live Demo starts on the cover — click Open desk to enter

Right Prospect

An AI-powered, signal-driven prospecting platform for B2B outbound sales.

Next.js Vercel Postgres pgvector Python JobSpy Embeddings

Right Prospect converts unstructured market signals into ranked account intelligence, starting with LinkedIn hiring data. It combines structured account filtering with semantic retrieval and cross-encoder reranking to match companies against a target ICP.

sensors

Signal Engine

Ingests LinkedIn hiring signals into Postgres with a decoupled, multi-source architecture.

neurology

Hybrid ICP Search

Combines structured filtering, vector retrieval, and cross-encoder reranking to surface ICP-aligned accounts.

database

Normalized Data Layer

Entity-level Postgres caching that eliminates redundant scraping and delivers low-latency queries.

Live & automated
June 2026
AI Builders Digest architecture: trigger, data collection, clean and prepare, AI processing, store and render, publish and display

How the pipeline works · tap to enlarge

AI Builders Digest

GitHub Actions Node.js DeepSeek Vercel Telegram Bot

I follow a focused set of AI builders to learn how they ship, but checking X, blogs, podcasts, and videos every morning was becoming its own routine. AI Builders Digest turns that habit into a timed pipeline: three times a week, it collects the last 72 hours of posts, publishes a bilingual (DE/EN) magazine automatically, and sends the new issue to me through a Telegram bot.

rss_feed

Extract

Collects fresh posts across X, YouTube, podcasts, and blogs, using RSS/Atom where available and keeping a rolling 72-hour window.

neurology

Transform

Posts are trimmed for launches, model names, and numbers, then passed to the DeepSeek API to generate the DE/EN issue.

autorenew

Load

GitHub Actions runs on Mon/Wed/Fri. Each issue is committed to Git, deployed on Vercel, and announced through a Telegram bot.

Live & iterating
May 2026

Public demo with sample data · opens best in a new tab

Job Match

Next.js Vercel Convex LLM Python Worker Apify

A personalized job-ranking and application-tracking platform for the Swiss market. It aggregates Swiss job postings, ranks each one against my CV and a structured requirements profile, tracks every application from first sight to final outcome, and records feedback to learn what works.

tune

Rank

Postings → CV and requirements profile → match score and priority.

fact_check

Track

Pipeline stages, status, shortlist, notes, and filters in one place.

insights

Learn

Outcome notes → feedback patterns → a clearer search strategy over time.

Live demo · agent loop in design
June 2026

Mobile demo with sample data · opens best on a phone

Scheduling Agent

Sparked by a conversation with a founder I met at a networking event: she wanted a calendar that actively manages her time, smart enough to know that a train ride can fit a low-priority meeting while a flight cannot.

Supabase Postgres + RLS Vercel LLM (BYO-key) Agent Loop Vanilla JS

A proactive scheduling agent that helps you run your day. Every time you open it, it scans your whole calendar and catches the things an ordinary calendar never warns you about, like a deadline with no time blocked for it, three meetings back to back, or an afternoon that is quietly overloaded. For each one it drafts a fix you can accept with a single tap, and every time you accept or pass it learns a little more about how you actually like to work and live.

visibility

Perceive

Six detectors watch for deadline risks, conflicts, follow-ups, and overloaded days.

psychology

Decide

It drafts a fix and proposes it. You stay in control and just confirm or pass.

trending_up

Learn

Every accept or pass tunes your preferences, so it fits your day better over time.

science

More explorations coming.

This space is reserved for upcoming AI experiments — agents, evaluation harnesses, retrieval systems, and product prototypes.