Back to portfolio

Case studies

Technical proof, not resume summaries.

Each case study separates status, role, architecture, reliability, tradeoffs, impact, stack, and evidence so hiring managers can evaluate the work quickly.

Mobile app QA build / private productEvidence page

Mi Nutricionista Mobile App

Patient-facing nutrition app that turns daily food tracking, chat, plans, appointments, and AI meal estimates into one mobile workflow.

Timeline: Built and captured through mobile QA flows in June 2026.

Team: Solo builder/operator.

Stack: React Native, Expo, TypeScript, Supabase architecture, AI meal estimation, mobile QA

Problem: Nutrition apps often split coaching, food logging, appointments, and progress across separate tools. The product needed a mobile surface where a patient can see goals, chat with a nutritionist, log meals, estimate macros, track progress, and keep the relationship active.

What I built:

  • Built a mobile patient experience with home dashboard, nutritionist chat, profile, plan, appointments, food diary, meal categories, and progress surfaces.
  • Implemented meal logging flows for text, photo, barcode-style entry points, manual macro review, and local AI nutrition estimates before save.
  • Designed the app around real patient behavior: daily goals, remaining calories, macro progress, check-ins, reminders, and low-friction actions.
  • Captured verified mobile screenshots from QA flows so the portfolio can show product proof without exposing production health data.

Architecture: Mobile app -> patient workflow screens -> nutrition coaching/chat surface -> meal diary and AI estimate flow -> private backend/product layer.

Reliability: Screenshots come from QA capture flows; private patient data, production credentials, and backend logs are not exposed.

Tradeoffs: Public proof shows product surfaces and flows instead of production data or private clinical workflows.

Impact: Adds a concrete mobile product proof point: not just SaaS dashboards, but a patient app with daily use loops and AI-assisted nutrition tracking.

Evidence: Mobile screenshot set: home, chat, dashboard, plan, appointments, diary, and AI estimate flows.

Public demo / private backendOpen live surface

Agent Bubble

Public AI agent interaction surface built to feel like a product, not a prompt box.

Timeline: Prototype-to-public demo sprint; current public surface is live.

Team: Solo builder.

Stack: Next.js, TypeScript, Vercel, AI workflow architecture

Problem: Agent demos often look like internal tools. The project needed a public surface that could communicate AI workflow value quickly to recruiters, clients, and collaborators.

What I built:

  • Built a public agent/chat product surface with responsive UI and production deployment.
  • Structured the experience around product proof instead of a generic chat interface.
  • Kept private agent actions, credentials, and production logs out of the public surface.

Architecture: Browser UI -> Next.js app -> private AI workflow/backend layer -> deployment on Vercel.

Reliability: Public route smoke-tested after deploy; private actions and credentials are intentionally not exposed.

Tradeoffs: Kept the public demo lightweight instead of exposing private backend controls or logs.

Impact: Provides a live AI-workflow proof link aligned with the portfolio positioning.

Evidence: Live demo + screenshot evidence page.

Private productionEvidence page

Titan Agent Fleet

Distributed AI operating environment across Windows, Linux, macOS, and GPU infrastructure.

Timeline: Built iteratively as the operating layer for real agent work across multiple machines.

Team: Solo builder/operator.

Stack: OpenClaw, LiteLLM, vLLM, Docker, systemd, Caddy, Tailscale, GPG vaults, Windows, Linux, macOS

Problem: AI work was fragmented across chats, tools, machines, credentials, codebases, browser tasks, and deployment environments. The system needed persistent memory, task delegation, model fallback, browser access, and production-grade execution discipline.

What I built:

  • Designed a multi-host agent operating model with memory/wiki context, tool permissions, sub-agent delegation, scheduled jobs, and verification rules.
  • Routed workloads by host capability: local coordination, GPU/model serving, media/video processing, browser automation, and infrastructure operations.
  • Integrated cloud model usage with local fallback through LiteLLM/vLLM and OpenAI-compatible endpoints.
  • Created SOPs for build checks, smoke tests, browser checks, endpoint validation, and incident-aware execution.
  • Used the fleet as a real execution team: briefing tasks, assigning workflows, reviewing output, correcting direction, testing results, and shipping artifacts.

Architecture: Telegram/user request -> OpenClaw orchestration -> tools/sub-agents -> local/cloud models -> verification gates -> deploy/report.

Reliability: Uses permissioned tools, vault-based secrets, build/lint/smoke gates, deployment checks, and explicit blocker reporting.

Tradeoffs: Private by design: host access, credentials, logs, and customer/business data cannot be public proof artifacts.

Impact: Reusable internal AI operations platform for building, debugging, deploying, and auditing software/business systems faster.

Evidence: Architecture notes + this case study + repeated deploy/build validation logs.

Private production architectureEvidence page

WhatsApp Business Agent

CRM and follow-up automation architecture for inbound business conversations.

Timeline: Built as CRM and follow-up automation inside a broader business operations stack.

Team: Solo builder/operator.

Stack: Next.js, Supabase, webhooks, WhatsApp architecture, Vercel, TypeScript

Problem: Manual WhatsApp follow-up creates missed leads, weak context, and repeated operator work. The system needed structured capture, classification, logging, and follow-up scheduling.

What I built:

  • Implemented inbound classification, contact/profile upsert, conversation logging, message logging, and follow-up scheduling.
  • Built webhook routes, inbox UI, simulated outbound fallback, cache fixes, local build checks, smoke tests, and production endpoint validation.
  • Kept private contact data, conversations, credentials, and business logs out of public proof surfaces.

Architecture: Inbound WhatsApp/webhook event -> classification -> Supabase contact/profile upsert -> conversation/message log -> follow-up schedule -> inbox UI.

Reliability: Local build checks, smoke tests, endpoint validation, no-store cache headers, and private-data boundaries.

Tradeoffs: Public case study describes architecture and scope without exposing real customers, phone numbers, private messages, or credentials.

Impact: Turns repeated WhatsApp follow-up into a trackable CRM workflow with clearer ownership, history, and next actions.

Evidence: Scoped case-study details + deployment/process validation notes.

Private production / public surfaceOpen live surface

nutricionista.ai

AI-enabled SaaS platform for nutrition professionals and patients.

Timeline: Multi-phase SaaS build with public surface and private role-based workflows.

Team: Founder-led / solo full-stack implementation.

Stack: Next.js, React, TypeScript, Supabase, Postgres, Vercel, Stripe architecture, shadcn/ui

Problem: Nutrition professionals need patient management, CRM, AI productivity, billing architecture, onboarding, and patient portal workflows in one system.

What I built:

  • Built admin dashboard, CRM, leads, patient workflows, appointments, resources, onboarding, integrations, and role-based surfaces.
  • Built patient portal PWA and SEO directory for nutritionist discovery.
  • Added AI feature foundations for meal planning, recipes, lab interpretation, growth, and progress analysis.
  • Implemented Supabase data layer and Vercel deployment workflows.
  • Shipped 35+ routes and 80+ components across admin, CRM, leads, patients, appointments, team roles, onboarding, resources, integrations, and patient portal surfaces.

Architecture: Next.js app -> Supabase/Postgres -> role-based surfaces -> Vercel deployment -> public marketing/SEO surface.

Reliability: Private health/patient workflows are not public; public surface is live. Claims are limited to route/component scope instead of unverified user/revenue metrics.

Tradeoffs: Public proof avoids exposing health data, admin workflows, billing internals, or private AI workflows.

Impact: Large SaaS surface with 35+ routes, 80+ components, multiple role-based workflows, patient portal, admin tooling, and AI feature foundations.

Evidence: Live public surface + screenshot evidence page + scoped case-study details.