Talles – AI-assisted coding, Design system, JavaScript

A cross-platform practice operating system for teachers, coaches, students, and athletes to log, teach, and manage real-world practice across yoga, strength training, school classes, and personal discipline. It combines class operations, personal practice, session recording, reflection, messaging bots, and AI-assisted insights around a durable execution ledger that preserves historical activity as practice catalogs and plans evolve.

• Designed and shipped the full-stack architecture (Next.js API, shared React Native/web client, Postgres, auth, media, bilingual EN/PT);
• Built a provider-agnostic LLM layer (xAI/Grok, Anthropic, OpenAI) with channel-aware prompts, model preference, and fallback on timeout, rate limit, and 5xx;
• Implemented a production STT pipeline (AssemblyAI and xAI) with job state, retries, webhooks/polling, diarization mapping, and language selection;
• Shipped continuous class-session tape: chunked recording at the STT cap, overlapping upload while the next clip records, pause/resume, and concatenated transcripts;
• Built voice roll-call: short audio → boosted STT on roster names → conservative name matching, with spoken aliases fed back into the next class;
• Implemented propose-then-confirm class notes: LLM drafts per-student comments from tape, recap, and hold-round clips as constrained JSON; teachers choose private vs shared; no silent auto-write;
• Built a clinician assistant that assembled session history and reference materials into LLM prompts and returned copy-ready comments and recommendations;
• Shipped Telegram and WhatsApp teacher bots with account linking, Redis-backed ingest, thread leases, retries, and outbound replies through the same chat completion path;
• Built the consumer “magic loop”: capture (text/voice/photo) → synchronous insight → async Grok Imagine image, stored on canonical media with quotas, idempotency, and prompt provenance;
• Designed the domain model for practices, plans, executions, comments, and linked media so AI outputs attach as first-class records, not chat leftovers;
• Integrated Cloudflare Images/R2, signed media URLs, and async generation jobs (STT, image) without blocking the user save path;
• Built class capture and student surfaces: email-less roster, claim, homework assignment, My Tasks, practice catalog, and personal sitting log;
• Wrote parser/unit tests for LLM JSON (drop unknown IDs, skip empty bodies, collapse duplicates) so model calls stay out of the test suite;
• Debugged live-class audio, STT format mismatches, and backgrounding so tape survived real rooms, not only demos.

A voice-controlled animal racing game for casual and betting-style players on mobile and desktop, where players select an animal and use its sound through a microphone to control its speed in real time while racing against AI bots. The product combines on-device machine learning, low-latency browser audio processing, and a Phaser-based game experience, with creator/streamer functionality, parent-host integration, encrypted game results, and supporting web surfaces for leagues and live broadcasts.

• Trained and exported an on-device animal sound classifier using YAMNet transfer learning and TensorFlow.js conversion;
• Built the end-to-end browser inference pipeline (AudioWorklet → Web Worker → TF.js WASM) for real-time voice gameplay;
• Achieved sub-100ms average inference latency on typical mobile devices;
• Implemented the Phaser race game: animal selection, mic warm-up, voice-driven physics, bots, results, and jackpot flows;
• Designed per-animal confidence/volume thresholds and difficulty systems (bot catch-up, guaranteed placements) for fair and monetized races;
• Integrated the game with a parent host via iframe postMessage and AES-GCM–encrypted result payloads;
• Built creator/streamer mode with webcam layout and infinite-race testing;
• Shipped Next.js game host and model demo surfaces for local validation of live mic inference;
• Built the leagues marketing site and transparent OBS stream overlay with live schedule / next-race countdown;
• Wrote Vitest coverage for the audio worklet and worker management layer;
• Tuned mobile layouts and performance (FPS throttling during model warm-up, worker keep-alive across scenes, transferable audio buffers).

A blockchain-integrated gaming product suite for crypto-native players, bettors, and operators, evolving from NFT minting and token staking into a live crypto betting and prediction-markets platform built around an on-chain racing game API. The product supported crypto betting, staking, rewards, NFT minting, automated on-chain payouts, multi-chain deployments across Ethereum, Arbitrum, Polygon, and Solana, and gamer-statistics services for thousands of concurrent web users.

• Led a ~10-person full-stack team shipping blockchain-integrated gaming products end to end;
• Designed and shipped a crypto betting platform (web, server, EVM contract) on a third-party on-chain game API serving thousands of users;
• Solved on-demand bet-pool creation with message queues and cross-system result polling/sync (third party, server, contract);
• Implemented automatic on-chain payouts after pool settlement;
• Built concurrent blockchain transaction creation from a single master wallet;
• Designed a stateless polling flow for bet-pool closing, result calculation, and crypto payouts via smart contract;
• Integrated Hardhat so API logic could be E2E-tested against real databases and local blockchain instances;
• Architected and led a PERC token staking/rewards platform shipped in under two months;
• Shipped an NFT minting path from web UI through chain, server, and spreadsheet operations;
• Architected and led a DAO voting platform with CMS-managed proposals;
• Architected xp.gg-style gamer stats: stateless APIs, heavy polling workers, and Bull queues handling thousands of third-party API calls and hundreds of DB queries per signup while respecting rate limits;
• Evolved the betting product into a prediction-markets platform and migrated Ethereum → Arbitrum → Polygon;
• Launched live-race betting with new APIs and a USDC betting web app;
• Rewrote the EVM contract as a Solana program and moved new APIs onto Solana;
• Coordinated third-party game API integrations and contractor delivery;
• Debugged and resolved production issues across web, services, and chain layers.

A decentralized storage blockchain network and product ecosystem supporting validators, miners, and storage providers, with web applications, CLI tools, and Web3 interfaces for network interaction. The ecosystem included a primary React-based application, multiple Next.js products, NFT minting and upload capabilities with Ethereum smart-contract integration, and client request flows supporting different consensus protocols across heterogeneous node types.

• Led a 12-person web frontend team through a full refactor; raised code-quality standards, coordinated delivery, prioritized work, and managed testers and contractors;
• Cleared the main defects on the core React/Redux/Saga/Web3 application (forms, crypto, wallet flows);
• Introduced Jest regression tests against live network APIs so UI tests were not coupled to on-chain state;
• Debugged Go services across validators, miners, storage providers, and CLI tools; isolated root causes and opened backend PRs;
• Wrote local Go scripts to exercise specific blockchain services outside full network setups;
• Improved core Go functions and supported backend teams with issue diagnosis;
• Managed five concurrent Next.js products while continuing hands-on development;
• Built and reviewed an NFT web studio for mint/upload, including Web3 and ETH smart-contract interaction;
• Implemented JavaScript request logic for multiple consensus protocols against heterogeneous network nodes;
• Enforced delivery strategy and quality standards so frontend products shipped on schedule.

An enterprise workplace access platform that uses smartphone-based digital credentials and BLE identity signals to replace physical badges and keycards. The platform enables employees, contractors, tenants, and visitors to access doors, elevators, and turnstiles while supporting property and security operators through web-based administration, visitor management, and real-time presence experiences.

• Prototyped a React visitor portal for lobby security and clerks, then migrated it to Gatsby so it could merge into the main dashboard;
• Built a high-performance visitor list with react-table, react-window, and react-virtualized so thousands of rows stayed responsive, including a custom class + React Context sticky header that matched the design spec;
• Mocked backend behavior with Netlify Functions on AWS Lambda so the prototype could ship without waiting on core APIs;
• Delivered the visitor prototype fast enough that design became the bottleneck; the PM and CEO commended the work in a company-wide meeting;
• Joined the core platform team on AWS TypeScript Node microservices: TDD, Express, TypeORM, MySQL, basic and bearer auth, independently scalable Lambdas with SNS/SQS between services;
• Co-built a Go visitor backend prototype (controllers, models, protobuf, gRPC) with Envoy + gRPC-web so the web client could call it; leadership kept Node because most of the team did not know Go;
• Shipped Gatsby dashboard work that loaded tens of thousands of rows on the client via virtualization for a property-access product handling millions of policy rows;
• Refactored Gatsby internals-level page creation, SSR, and code splitting and cleared team questions on those internals;
• Reduced the core-platform backlog, then took on main dashboard/backend tasks, then returned full-time to the renewed visitor MVP as that work resumed;
• Wrote Jest unit tests and Storybook stories; used CircleCI, Datadog, Jira, GitHub, and Notion in day-to-day delivery;
• Built a real-time presence demo: vanilla HTML/JS client, Socket.io, Heroku Node webhook into the main backend, showing identity and photo when LBE sensors detected a person;
• Collaborated with PMs and other teams on extra app demos/prototypes; received strong manager reviews and helped absorb overload so new-product work could restart (and helped open the door for four more Turing hires).