I take AI from research to products that real businesses run on.

Problem

Spinal anaesthesia is traditionally a blind procedure: the clinician finds the right gap between two lumbar vertebrae by feeling the patient's back (manual palpation), then judges the needle's entry point and angle by touch alone. Roughly 1 in 3 first attempts fails — repeat punctures, complications, wasted clinical time. Ultrasound is the recommended remedy, but reading spinal ultrasound is genuinely difficult. uSINE is the AI guidance layer in between: it identifies spinal landmarks in real time and guides needle placement.

Approach

• Moves with the clinician: adoption drove the design — clinician interviews set the direction, the UI was co-designed with them and a professional UX team, and the tracking algorithm was fully redesigned, even under time and cost pressure, so clinicians could scan the way they already do.
• Engineered for the procedure room, not the lab: runs alongside the ultrasound machines hospitals already own, with deep algorithm optimization delivering real-time performance on standard off-the-shelf CPU laptops — lower cost, portable, nothing to replace, and nothing exotic to install or maintain.
• Built to be audited: extensive testing and validation — including clinical validation across 200+ patients — thorough documentation, and Git-based traceability that keeps every change auditable on the path to multi-market regulatory approval.

Tech stack

Real-time computer vision on off-the-shelf CPU laptops — OpenCV, Cython, ONNX, PyQt

My role

CTO & co-founder — led the full prototype-to-production rebuild across algorithms, backend, and UI.

Problem

A hair-care chain wanted every customer to get the same quality of scalp consultation at every outlet — hard when diagnosis depends on a handful of scarce trichologists. Objective, quantitative scalp diagnosis — grounded in hair and follicle density — is impractical to do manually, so in practice it was either skipped or bottlenecked on those few experts.

Approach

• Explainable hybrid AI: deep learning detects only what a human can verify with their own eyes — hair strands, follicles, dandruff, pimples — and a rule-based expert system turns those observations into the diagnosis. Trust is what drives adoption: staff and customers can always see why.
• Improves with every scan: a large labeled dataset was out of reach upfront, so the team shipped a workable product first and let production data do the rest — each new scan feeds an MLOps loop of anonymization, cleaning, model-in-the-loop pre-labeling, human review, retraining, benchmarking, and release.
• Built to run a business: multi-tenant architecture with role-based access control, hybrid subscription plus pay-per-use billing, and a CRM and analytics dashboard that outlet staff use in daily operations — delivered as a PWA: one app on every outlet PC and iPad.

Tech stack

Multi-tenant SaaS (FastAPI, PostgreSQL), React PWA client, PyTorch → ONNX, CVAT + W&B ML pipeline, Stripe & Twilio, Docker + CI/CD

My role

CEO & co-founder — took it from the first business conversation to production and 15 months of operations, leading a 5-person team.

The concept

The name fuses Organism and robot: a companion robot exploring the middle ground between living things and machines. It watches over a plant's well-being and gives it a mobile, playful presence in the home — art and engineering in one build.

Approach

• Agent-native development: the repo is structured for AI coding agents as day-to-day contributors — scoped instructions, path-scoped rules, and hooks that run a fail-fast quality gate after every edit, holding agent- and hand-written code to the same typed, tested, documented bar as my shipped products.
• Strict separation of concerns: hardware drivers, assembly logic, the API service, and the web dashboard are cleanly layered, and backend and frontend share only one HTTP contract — each side develops and tests independently.
• Cost-conscious product architecture: designed from day one to a consumer-viable cost point — standard off-the-shelf parts wherever practical, down to bearings and screws, with deliberate trade-offs across the hardware, electronics, and software stack.

Tech stack

Raspberry Pi · Python (Falcon ASGI) · SQLite · Vue 3 + Quasar · Autodesk Fusion (CAD) · 3D printing

My role

Solo builder — mechanical design and 3D printing, electronics, sensing and control, backend, and app: every layer is one person's work.