Systems built.
Problems solved.

What it is

A SaaS platform that builds complete business websites, for any type of business, through a multi-agent AI orchestration system. The platform coordinates 23+ specialised AI agents across 9 functional departments; each agent has a defined role, domain, and personality, and together they run 31 automated workflows end to end, from brief to published site.

The interesting problem

Building a demo with 5 agents is straightforward. Building a system where 23+ agents coordinate reliably in production is a different problem entirely. It requires idempotent state management, queue-based processing that avoids lock contention under concurrency, self-improvement loops gated by staging environments, and graceful degradation when upstream models behave unpredictably.

The failure modes in multi-agent systems are not the failure modes people expect. Agents do not usually fail loudly. They fail by producing plausible output from a wrong premise, and that error propagates silently through every downstream agent. Detecting that propagation before it reaches the user is most of the engineering.

Result

In production, running live. Average end-to-end production cost of $0.08 per finished business website.

Technologies (selected)

Python, TypeScript, Node.js, LangChain, LangGraph, PostgreSQL, vector databases, multiple LLM providers, event-driven background processing, real-time WebSockets.

What it is

A complete architecture design for an algorithmic trading system targeting crypto derivatives markets. The system is built around a real-time data pipeline that moves from binary market feeds through a time-series database to feature engineering and ML inference, with a convergence-scoring framework combining 6 parallel market signals.

The interesting problem

Latency budgets of 10 to 50 microseconds for the hot path require architectural decisions that most software engineers never face: zero-copy deserialisation, bare-metal cloud deployment, and an ML inference layer that cannot afford a garbage collector pause.

The convergence framework (liquidation cascade, cross-exchange divergence, order book imbalance, funding rate, VPIN, gamma exposure) is designed to reduce false signals while preserving sensitivity to genuine market dislocations. Any single signal produces noise. The engineering problem is scoring agreement across six signals fast enough that the answer still matters when it arrives.

Technologies (selected)

Rust (hot path), Python (feature engineering and ML), LightGBM, time-series database, real-time binary feed parsing.

What it is

A full-stack platform for the real estate market that uses generative AI to automate market research; property content generation across text, image, and video; intelligent property matching; and multi-channel publishing across 4+ platforms.

The interesting problem

Real estate content is highly repetitive at the surface. Every listing has a description, photos, and a target audience. But it requires contextual intelligence at the detail level: local market dynamics, comparable properties, target buyer persona.

The system needed to be opinionated enough to produce good output by default and flexible enough to handle the genuine variation across property types and markets. Those two requirements pull in opposite directions, and most content generation systems resolve the tension by producing output that is technically correct and commercially useless.

Technologies (selected)

Next.js, React, Node.js, PostgreSQL, multiple generative AI providers (text, image, video), WhatsApp Business API, event-driven background processing.

What it is

A multi-modal biomarker fusion engine for early, non-invasive disease risk detection. The system ingests structured physiological signals from already-certified sensors (ECG, glucose, dermatological imaging) and correlates them across domains, instead of scoring each in isolation, to surface risk patterns that single-modality tools miss.

The interesting problem

Cardiovascular, metabolic, and dermatological signals are still captured and read in isolation. The early warning patterns that only emerge across these domains are the ones no single-modality test is designed to detect.

The fusion architecture has to handle signals with fundamentally different data structures, sampling frequencies, and clinical significance levels, and produce a composite risk score that a clinician can interpret and act on. A score nobody trusts is worse than no score, so interpretability is a hard constraint rather than a feature.

Current status

TRL 4. End-to-end feasibility demonstrated on synthetic datasets modelled on published biomarker distributions. Synthetic data proves the pipeline, not the signal, which is exactly why the next phase runs on real-world data.

That phase benchmarks the cardiovascular and metabolic fusion layers against public datasets, PhysioNet first and MIMIC-IV as a stress test, acknowledging its ICU population is not the target cohort. The dermatological modality, for which no paired public dataset exists, will be validated through a clinical pilot, with the search for a Portuguese clinical partner running in parallel.

What it is

I contributed to commercially released digital products, from concept through production deployment, under non-disclosure agreements. Full product lifecycle: architecture design, development, testing, deployment, and operation with real users at scale.

What can be shared

I worked in multidisciplinary teams, shipped critical features under production constraints, and made direct, measurable contributions to commercial outcomes. The products themselves stay under NDA, and I can go deeper on the engineering problems in a call.