01 · currentbuilding now
HostHive
Own project · marketplace
A two-sided marketplace connecting companies with professional event and trade show hosts.
Read more about HostHive →AI engineer · researcher · creator
Karmel Salah
I build machine learning and LLM systems that run in production, not in notebooks.
Current projects
What I am building now, as my own projects. Open any card for my role and the details.
- 01 · current project
HostHive
Own project · marketplaceMy role
Founder and product owner. I defined the marketplace, both sides of the experience (companies and hosts), and the roadmap, and I lead the build of the website and the mobile app.
What it is
A two-sided marketplace connecting companies with professional event and trade show hosts: a Next.js website and an Expo mobile app on Supabase.
Approach
A Next.js website and an Expo mobile app, both on Supabase.
Stack
- Next.js
- Expo
- Supabase
Status
In progress.
Links
Link pendingFull page →/work/hosthive
- 02 · currentbuilding now
Go
Own project · trip planningA hands-free voice co-pilot for drivers and riders, with Gobie as its voice companion.
Read more about Go →02 · current projectGo
Own project · trip planningMy role
Founder and lead engineer. I came up with the idea of a co-pilot you can talk back to mid-ride, designed the agent's behaviour, and lead the backend, the voice agent and the mobile app.
What it is
A hands-free, voice-conversational navigation co-pilot for drivers and riders wearing headphones, with its own turn-by-turn navigation. The plan adds trip planning on local routes, with Gobie as the voice companion.
Approach
A FastAPI and WebSockets backend, a Gemini agent that knows the live GPS position and the route, and an Expo mobile app.
Stack
- Voice agents
- Gemini
- FastAPI
- WebSockets
- Expo
Status
In progress. Private repository.
Links
Private repositoryFull page →/work/go-gobie
- 03 · currentdemo mode
Go Mini
Own project · route rewardsA small version of Go: add the route you prefer, earn points and get coupon rewards.
Read more about Go Mini →03 · current projectGo Mini
Own project · route rewardsMy role
Founder. I scoped Go Mini as a lighter first release of Go, focused on sharing favourite routes and rewarding people with points and coupons.
What it is
You open the app, add the route you prefer, earn points and get coupon rewards.
Approach
It runs alongside the full Go (trip planning with Gobie). Coupons stay marked "demo" until a legal check.
Stack
Details coming soon.
Status
Being built in demo mode. Android preview expected around 9 October 2026.
Links
Links pendingFull page →/work/go-mini
- 04 · currentdevnet test
QSwapie
Own project · Solana, devnet testA virtual queue for busy cafés, restaurants and shops, where you can offer to swap places.
Read more about QSwapie →04 · current projectQSwapie
Own project · Solana, devnet testMy role
Founder. I invented the swap-your-place idea and designed how a fair, non-custodial swap and payment settle in one transaction; I lead the on-chain program and the web app.
What it is
Customers scan a rotating QR code to take a number and can offer SOL to swap places; the swap and payment settle in one transaction. Built on Solana: non-custodial, devnet test.
Approach
An on-chain program (Anchor) and a Next.js web app. The swap and the payment settle together in one non-custodial transaction.
Stack
- Solana
- Anchor
- Next.js
Status
Web app live at qswapie.de. Payments are non-custodial and on Solana devnet (a test network): no mainnet, no live money.
Links
/work/qswapie
- 05 · currentbuilding now
Hands Up
Own project · accessibility AIReal-time sign language to spoken language, on the web and on the phone.
Read more about Hands Up →05 · current projectHands Up
Own project · accessibility AIMy role
Founder. I started Hands Up to make conversation between Deaf signers and hearing people easier, and I lead its research, privacy-first design and the on-device recognition approach.
What it is
Real-time sign language to spoken language on the web and phone. Version 1 is ASL to English, with Arabic and Hebrew later. Private by design (on-device), built with the Deaf community.
Approach
On-device sign recognition, so conversations stay private, shaped together with Deaf signers.
Stack
- Computer vision
- On-device ML
- Accessibility
Status
In progress.
Links
/work/hands-up
Previous experience
Roles and projects from before. Each card says what I did there.
- 2022 →01 · previous
Pic-Time
AI Album Designer
AI EngineerA real-time feature that turns a photographer's full shoot into a finished album.
Delivered a 10x improvement in inference throughput in automated album design · dataset curation across 50,000+ galleries
Read more about AI Album Designer →01 · previous experience · Pic-TimeAI Album Designer
AI Engineer · 04/2022 to 09/2026My role
Designed and built the AI Album Designer as part of the AI team: the multi-model pipeline (face recognition and clustering, automated photo selection, saliency-aware smart cropping) and the inference work that lets albums generate while the photographer waits. Also shipped face-based gallery access and keyword- or face-based photo search. It started as an internship project and grew into a core AI team and product pillar.
What it is
Photographers upload a full shoot and get a finished album on demand. Alongside it: face-based gallery access and photo search on embedding similarity, served as REST APIs.
Approach
Multi-stage inference tuned for speed; near-duplicate rejection, album-wide consistency and keeping faces intact under aggressive crops. Photo search uses embedding-based vector similarity behind REST APIs.
Stack
- Python
- FastAPI
- Docker
- Kubernetes
- Multimodal transformers
- Embeddings · vector search
Status
Shipped and in production.
Delivered a 10x improvement in inference throughput in automated album design · dataset curation across 50,000+ galleries
Links
/work/pic-time-album-designer
- 2021 to 202202 · previous
Morphysm · for Deutsche Oper Berlin
GPT-3 opera assistant
AI EngineerOne of the first GPT-3 conversational products in Berlin's cultural sector, shipped as a Telegram bot.
27,000+ monthly users, achieving an 80% reduction in manual support volume
Read more about GPT-3 opera assistant →02 · previous experience · Morphysm · for Deutsche Oper BerlinGPT-3 opera assistant
AI Engineer · 07/2021 to 03/2022My role
Built and deployed the assistant: designed the conversation flows, the prompt structure and the fallback handling, and built the React interface. Wrote the repository docs and video tutorials for onboarding.
What it is
A GPT-3 conversational assistant for Deutsche Oper Berlin, shipped as a Telegram bot that answered real audience questions.
Approach
Structured prompts and conversation flows, with fallback routines for the questions the model should not answer itself.
Stack
- GPT-3
- Telegram bot
- Prompt design
- React
Status
Shipped.
27,000+ monthly users, achieving an 80% reduction in manual support volume
Links
/work/morphysm-deutsche-oper
- 202103 · previous
aLifeForms
Generative AI gallery
Software engineer and AI researcher (freelance)Experimental AI work for an artist: an online gallery of interactive, generative works.
Read more about Generative AI gallery →03 · previous experience · aLifeFormsGenerative AI gallery
Software engineer and AI researcher (freelance) · 04/2021 to 07/2021My role
Co-built an online AI gallery of interactive, generative works that combine computer vision and audio, and ran face-decomposition experiments.
What it is
Interactive, generative artworks that combine computer vision and audio, for an artist's online gallery.
Approach
Real-time generative pipelines, packaged in containers for digital art installations.
Stack
- Computer vision
- Audio
- Generative pipelines
Status
Exhibited work.
Links
/work/alifeforms-generative-art
- 2020 to 202104 · previous
NoCompany
Monster In The Middle
Software Engineer (freelance)An open-source web app that teaches networking and man-in-the-middle attacks visually.
Read more about Monster In The Middle →04 · previous experience · NoCompanyMonster In The Middle
Software Engineer (freelance) · 12/2020 to 04/2021My role
I built the app end to end, including Wi-Fi connectivity simulations.
What it is
An educational web app, supported by the ISC Project, that makes network attacks visible.
Approach
React and Material UI front end, Node.js back end.
Stack
- React
- Node.js
- Material UI
Status
Open source.
Links
/work/nocompany-mitm
- 202005 · previous
Gaza Sky Geeks
Academy and HR web apps
Full-stack web developer internAn application web app for Gaza Sky Geeks' coding academy, and an Arabic HR management app.
Read more about Academy and HR web apps →05 · previous experience · Gaza Sky GeeksAcademy and HR web apps
Full-stack web developer intern · 01/2020 to 10/2020My role
Built both apps end to end, from requirements to deployment.
What it is
An application web app for the GSG coding academy and an Arabic HR management app.
Approach
React, Node.js and Material UI.
Stack
- React
- Node.js
- Material UI
Status
Delivered.
Links
/work/gaza-sky-geeks
- 201806 · previous
EF English First
Online English-learning programme
Co-founder · CTO and COOA platform that connected English students with teachers.
Read more about Online English-learning programme →06 · previous experience · EF English FirstOnline English-learning programme
Co-founder · CTO and COO · 04/2018 to 11/2018My role
I co-founded it and ran both technology (CTO) and operations (COO).
What it is
An online English-learning platform connecting students and teachers. The authoritative CV calls it the "Within 6 Months" programme at EF.
Approach
Details coming soon.
Stack
Details coming soon.
Status
2018.
Links
/work/ef-english-first
- independent07 · previous
Independent research
AI job-risk and emotion study
ResearcherEnd-to-end NLP research on Reddit data about how people perceive AI job risk.
Read more about AI job-risk and emotion study →07 · previous experience · Independent researchAI job-risk and emotion study
ResearcherMy role
Details coming soon.
What it is
Topic modelling, staged LLM-assisted labelling (Grok, then GPT-4o-mini, then Mistral) and fine-tuned classifiers for fear, anxiety and hope.
Approach
Topic modelling, staged LLM-assisted labelling (Grok, then GPT-4o-mini, then Mistral) and fine-tuned classifiers for fear, anxiety and hope.
Stack
- NLP
- LLM fine-tuning
- BERTopic
Status
Public repository.
Links
/work/ai-job-risk-research
Research and writing
- R1
Kidney stone detection
Finding kidney stones and their location in CT scans with deep learning and CLIP-based approaches.
- R2
Solving SAT with a parallel genetic algorithm
A high-performance computing project.
- R3
Auto-complete with LDA and a character LSTM
Topic-aware autocomplete from an LDA topic model and a character-level LSTM.
Blog
First notes coming.
Build notes and research write-ups will appear here, dated, once they are real.
Let's talk
- Emailhello@karmelsalah.com
- GitHubgithub.com/karmelyoei
- About meRead my story →