AI apps & agents
LLM features that are useful in production: RAG over your own documents with citations, assistants that call tools, AI-written reports and analysis.
Open to roles across Europe · On-site, hybrid or remote
I build systems that automate operations.
AI & Software Engineer · Co-founder & CTO at SolverCam (NVIDIA Inception). I turn manual, messy work into software: AI apps, integrations, full-stack products and edge AI.
01 — What I build
Four kinds of systems, one goal: less manual work.
Automation & integrations
Connecting the tools a business already uses — payments, shipping, ads, ERPs, devices — so data flows on its own and people stop copying spreadsheets.
Full-stack products
From database to UI: multi-tenant SaaS, admin panels, online stores and mobile apps, built to be maintained — typed, tested and documented.
Edge AI & computer vision
Models that run where the data is: NVIDIA Jetson devices processing video on-site, with only structured data going to the cloud.
02 — Work, explained
Don't take my word for it. Press the buttons.
Each project below has a small live demo that shows the idea behind the real system.
JurisPI AI · SaaS
Answers from your own documents — with sources.
Law firms lose hours searching case files. Generic chatbots invent answers, which is unacceptable in legal work.
I built a multi-tenant SaaS where each firm uploads its documents. Text is split into chunks, embedded in pgvector, and every answer cites the exact passage it came from.
- Stack Next.js · Prisma · PostgreSQL + pgvector · BullMQ · Claude/OpenAI
- Key decision Every query is isolated by organization; the AI provider sits behind one interface so models can be swapped.
Ask the firm's documents:
embed question→
search vectors→
rank chunks→
answer + cite
Meta Ads AI Dashboard
An analyst that reads the ad account for you.
Agencies managing many ad accounts spend hours in spreadsheets to find which campaign is burning budget.
The tool pulls data from the Meta Graph API, a rules engine flags problems (low CTR, ad fatigue, days without results, budget pacing), and Claude explains them and writes the executive report.
- Stack React 19 · Meta Graph API · Claude · Vercel Functions · jsPDF
- Key decision Deterministic rules find the issues; the LLM only explains them. Cheaper, faster and auditable.
| Campaign | CTR | Freq. | Spend | Status |
|---|---|---|---|---|
| Retargeting · BR | 2.4% | 1.8 | €410 | — |
| Prospecting · Wide | 0.4% | 1.2 | €980 | — |
| Lookalike 1% | 1.1% | 5.6 | €640 | — |
| Brand · Search | 3.1% | 2.0 | €220 | — |
Production e-commerce
Selling the last item only once.
With instant payments (Pix), minutes pass between "I'll pay" and "paid". Decrementing stock only at the end sells the same piece twice.
Stock is reserved when the order is created and released automatically if payment doesn't arrive in 45 minutes. Payment webhooks are verified by signature; secrets never touch the browser.
- Stack React · Supabase (Postgres + Edge Functions) · Mercado Pago · Melhor Envio
- Key decision The public key can only read published products; every write goes through a server function.
Stock of "Linen dress, M": 1
- Order created
- Stock reserved · 45 min
- Payment confirmed (signed webhook)
- Shipping label requested
Engineering operations
A team that ships with AI agents — safely.
AI coding agents are fast, but without structure they create chaos: unreviewed changes and lost context.
At SolverCam I set up an operator → executor workflow on GitHub: tasks go into a queue, agents pick them up and open changes, and nothing merges without review. Decisions live in 30+ ADRs so humans and agents share the same context.
- Stack GitHub · Shell / PowerShell · Claude Code · ADRs
- Key decision The written decision record is the source of truth — for people and for agents.
Queue
Agent working
Human review
Merged
SolverCam · NVIDIA Inception
Video in, only numbers out.
Stores have cameras but no data. Streaming video to the cloud is expensive and a privacy problem.
A NVIDIA Jetson next to the store's recorder detects and tracks people on-site with DeepStream + TensorRT. Only events go to the cloud, through an offline-safe queue. In production: 7 cameras on one device, 110k+ events in 11 days.
- Stack Jetson Orin Nano · DeepStream · TensorRT · FastAPI · PostgreSQL · Next.js
- Key decision Process where the data is born. Privacy, bandwidth and cost all improve at once.
Entries0
Sent to cloud0 KB
Images leaving store0
03 — How I work
From a messy process to a measured system.
01
Map the work
Watch how the task is done today, where time is lost and what "done" means in numbers.
02
Decide in writing
Architecture Decision Records: options, trade-offs and why. Cheap to change on paper, expensive in code.
03
Build the thin slice
One end-to-end flow working in production first, then widen. AI agents speed up the build; reviews keep quality.
04
Measure, then scale
Metrics on day one. A milestone is a question answered with real data, not code delivered.
04 — Toolbox
Tools I use in production.
05 — Contact
Have a manual process that should be software?
I'm open to Software, AI and Automation Engineer roles anywhere in Europe — on-site, hybrid or remote.