Enclave is a Wallet built to embrace non-blockchain users. It combines cutting edge technologies like Abstract Accounts, WebAuthn and Gasless transactions to make onboarding as smooth as a traditional web application — paired with a full on-chain activity explorer. Working with a small team, I was responsible for the product vision, usability, and development of the whole wallet frontend, contributed to the Smart Contracts, and led the system design of the indexer and API behind the explorer. One example of that indexing work: handling chain reorgs. Every block was stored keyed by its own hash and its parent's, so detecting a reorg meant walking back parent hashes until the indexer found one it already had — the fork point. From there, the orphaned blocks and the balances derived from them were invalidated, and the canonical chain was replayed forward through idempotent handlers, so reprocessing never double-counted a transaction.

Neon is the leading wallet in the Neo ecosystem, with over $1 billion in traded volume. I was responsible for architecting its mobile version and later contributed to the desktop app. During my time on the project, I tackled key challenges such as supporting multiple blockchain networks, managing multiple accounts simultaneously, implementing WalletConnect integration, and developing the protocol for network interaction, along with several other critical integrations.

Louis Dreyfus Company, one of the largest commodity traders in the world, commissioned a 'Safety, Health, and Environment' management platform for use across all its global units — each with its own local regulations, forms, and reporting requirements. I owned it end to end: architecture, development, and integration with Azure Active Directory for authentication and user management. The core challenge was translating that sprawl of per-unit requirements into one flexible application instead of one branch per country, so I modeled forms, fields, and workflows as configurable data instead of hardcoded screens — letting a compliance team in a new unit stand up its own SHE process by configuring the platform, not by requesting a code change and a deploy.

Jamef, the largest shipping company in Brazil, needed a new dashboard for customers to track delivery data, but the existing one suffered from severe performance problems. My scope started as frontend-only, but tracing the slowness back to its source pulled me into the data layer. I introduced a caching layer and a server-side pagination policy for the result set, taking a query that took 8 seconds down to a few milliseconds. I delivered a complex dashboard with several customized graphs and contributed to restructuring the underlying central system.

I was the lead engineer who built iTrack Brasil from the ground up, owning its system design and scaling it into a B2B delivery platform integrating multiple systems, with nearly 60,000 couriers and over 50 million invoices processed across 2,000 registered companies — growth that led to its acquisition by MadeiraMadeira in 2021. One example of that scaling work: as invoice volume and courier position updates grew, the database hit a write bottleneck, surfaced by a spike at the end of a fiscal month. The system stayed up, but latency and AWS costs climbed, so I redesigned the write path at its source — courier position pings now land as Parquet files on S3, backed by a Redis cache holding each courier's last known position for fast reads, while invoice processing moved behind an SNS/SQS pipeline that absorbs bursts asynchronously.

Apptite was a food delivery app for iOS, Android and the web. It gained recognition with acceleration by '500 Startups' and media coverage that established it as an important platform in the artisanal food market. I was the main engineer responsible for the platform from its initial planning and structuring through scaling it as it grew. One example of that scaling work: the dish recommendation engine. The map was discretized into cells, each holding the list of stores that served it, updated in batch whenever delivery areas changed — so the geographic lookup itself was just a cached read. Personalization ran on top of that already-filtered set, combining complementary signals in parallel (repurchase history, regional popularity, among others), with the heavier estimates — including a store-by-cell matrix — precomputed offline.