Octopus Lab
AI on the edge — engineering studio
Tailored AI for complex, non‑standard problems.
A deep-tech studio led by an engineer and a physicist. Optimization, computer vision, and LLM automation that ship — measured against your KPIs, your latency budget, your hardware.
Built for
Series A teams & industrial R&D
Years applied R&D
Founders, full-stack
Capability lines
KPI-accountable engagements
What we build
Four capability lines
- 01
Mathematical optimization
Routing, scheduling, allocation, network flows. Solvers tuned to your data and infeasibility modes.
- 02
Computer vision
Detection, classification, tracking, OCR. Trained on your data, tuned for the GPU, NPU, or microcontroller it has to run on.
- 03
LLM automation
Workflow agents, document pipelines, RAG, evaluator harnesses. Measurable accuracy gates, not vibes.
- 04
Edge AI & embedded
Models that run where the data is born — on the gateway, the camera, the controller. Power, latency, thermal as first-class.
Proof points
Numbers we stand behind
Throughput on a vision line.
Detection model re-architected and quantized for the existing edge GPUs.
Optimization runtime.
Routing solver brought below dispatch SLA, with feasibility guarantees on bad-data days.
From brief to first KPI in production.
Hand-wave deliverables.
Code you own, tests you can run, a metric you can audit.
How we work
Four stages, no surprises
Discovery
One technical conversation. We agree the KPI, scope the risk, and decide whether a PoC is warranted.
Scientific PoC
A focused experiment with a measurable bar. We report pass or fail — no sunken-cost continuation.
Production build
Code, tests, observability. Built to the same bar you would hold your own engineers to.
Handover
Your team owns it. We document, transfer, and stay reachable — without a support retainer in the way.
Three commitments
- Honest negatives — we tell you when the experiment fails.
- You own the IP — code, weights, notebooks, day one.
- No lock-in — boring infrastructure on purpose.
Why Octopus Lab
An unusual kind of studio
Not a consultancy
The people who scope the work write the code. There is no handoff to a delivery team with a different understanding of the problem.
Not a platform vendor
We don't sell you a product to fit the problem to. The solution is built around your data, your hardware, and your KPI.
Engineer-led, physics-honest
Comfortable with messy data and real hardware. We know what the numbers mean — and what they cannot tell you.
The team
Two founders, full-stack
Carlo Perassi
Co-founder · Engineer
Software engineer with a background in distributed systems and mathematical optimization. Leads technical architecture and production delivery.
LinkedIn
Giacomo Leonzi
Co-founder · Physicist
Computational physicist and ML researcher. Leads model design, scientific validation, and edge deployment strategies.
LinkedInVoice
“They behaved like an internal team from week one.”
Partners & alliances
Academic ties: Politecnico di Torino · Università degli Studi di Torino
FAQ
Common questions
How do engagements start?
A single technical conversation — usually twenty minutes. We discuss your KPI, the data you have, and the hardware or infrastructure it needs to run on. From there we either propose a PoC scope or tell you we are not the right fit.
Fixed-price or T&M?
Both, depending on scope clarity. Discovery and PoC phases are typically fixed-price with a defined deliverable. Production builds are often T&M with a capped envelope, so you see the burn and can adjust.
Who owns the code and models?
You do. All code, trained weights, notebooks, and documentation are transferred to you at the end of each phase. We retain no licence rights and place no usage restrictions on the work.
Will you sign our NDA / MSA / DPA?
Yes. Standard confidentiality, data-processing, and master service agreements are straightforward for us. We are registered in Italy under Kiwifarm Srl, so Italian and EU law governs by default, though cross-border terms are workable.
Can you run inside our infra?
Yes. We regularly work inside customer VPCs, air-gapped environments, and on-prem GPU clusters. Bring your constraints early and we will design around them, not around our own tooling preferences.
Do you take research-only engagements?
Selectively. If the problem is a well-defined scientific PoC with a measurable success criterion, yes. Open-ended research without a production path is not our default — but we are comfortable being honest if the answer is not yet known.
Get in touch
Tell us the constraint. We'll tell you the shape.
Twenty minutes is enough to know whether we are the right bench. Bring the KPI, the data, and the hardware target.
Engineer-to-engineer. No sales call. Response within 48 hours.