01Lead developer
Property marketplace & agency CRM
A public property marketplace plus a multi-tenant CRM where agencies manage listings, leads and agents. I own it end to end — schema, API, authorisation, frontend and the AI surface.
- Agencies can trust who sees what. I designed the multi-tenant permission model so access reflects the team as it is today rather than as it was when someone last signed in.
- Buyers search by describing what they want. Natural-language property search in Albanian and English, with a deterministic fallback so results keep coming even when the model does not.
- Agents stopped writing listings by hand. Listing copy is generated from a property’s own photos and data, bounded so it stays factual.
- One product, three markets — Albanian, English and German throughout, including transactional email, with backend errors returned as translation keys so nothing ships untranslated.
- Releases stopped being nervous: a Vitest and Playwright suite with CI running end to end against a live database.
The AI surface
- What it does
- Natural-language search across listings, and descriptions written from a property’s own photos and data — the model reads the images, the copy comes back ready to edit.
- What it solves
- Buyers could not say what they wanted through dropdowns, and every agent wrote every listing by hand. Both were the reason stock sat unseen.
- What it taught me
- Treat a model as a dependency that will fail. Every call sits behind one layer with a deterministic path underneath, so a slow or wrong answer narrows the search instead of emptying it.
- React 19
- Vite
- tRPC
- NestJS 11
- Drizzle ORM
- PostgreSQL 18
- PostGIS
- Redis
- Nx
- Playwright
02Core contributor, both repositories
Consultancy CRM & operations platform
AI-powered internal platform
Internal operations for a consultancy: clients become projects, projects get consultants, consultants log time, time becomes reports, reports become invoices. An admin console and a consultant portal from one build.
- Consultant CVs are generated, not written. A consultant’s profile becomes a branded, client-ready document, tailored to the project it is being sent for, and reviewed by a human before it goes out.
- Onboarding went from an afternoon to a minute: an uploaded CV is read and turned into a structured profile — skills, experience and languages — for review rather than retyping.
- Finding the right consultant became a search rather than a conversation, with filtering and matching across skills, seniority and languages, and ranking a project manager can see the reasons behind.
- Clients get their reports without anyone remembering to send them — scheduled delivery in PDF, Excel or CSV, reliably and only once.
- No vendor lock-in on AI spend: one integration layer across four LLM providers, so changing provider is a configuration change rather than a rewrite.
The AI surface
- What it does
- Filtering and matching across skills, seniority and languages, and a client-ready CV generated from the consultant profile that match returns.
- What it solves
- Finding the right consultant meant asking colleagues who they knew, and every CV was retyped for each client it was sent to.
- What it taught me
- Generated documents need a human before they leave, and a provider you can change your mind about. One layer across four providers, so switching is configuration rather than a rewrite.
- NestJS 11
- Prisma 6
- PostgreSQL
- React 18
- TypeScript
- Storybook
03Team project
Medical imaging analysis platform
AI-assisted diagnostics
AI-assisted diagnostics for clinicians: DICOM study upload, patient records, and structured prompting against a medical vision-language model so findings come back parsed rather than as prose.
- Structured, modality- and body-part-specific prompting, so the model returns fields the interface can render instead of a paragraph a clinician has to re-read.
- Localised for Albanian-speaking clinicians, with JWT-protected patient records behind it.
- NestJS
- React
- PostgreSQL
- Nx
- Python
- Medical VLM
- DICOM
04Core backend & end-to-end flow
Influencer campaign platform
Brand & creator campaigns with an AI video studio
Brands and creators matched, briefed and invoiced in one place, with an AI video studio where a roster of AI presenters delivers a brand’s script — so a campaign can ship without a shoot.
- I built the core backend and the end-to-end campaign journey, from creation and editing through to delivery and invoicing.
- Brought the AI presenters to life on the front end, so the roster plays as video rather than sitting as stills.
- Next.js
- Drizzle
- PostgreSQL
- AI video
- S3
- Resend