The Challenge
Building production platforms across three distinct domains — healthcare, Web3, and AI — each with different tech stacks, regulatory requirements, and performance constraints. The challenge was applying consistent production discipline across diverse architectures while adapting to each domain's unique requirements.
Projects are described with technical specifics but anonymized per NDA agreements.
Project 1: Healthcare Platform
A mobile-first healthcare application serving 10k+ users across 4 Flutter apps (iOS, Android, 2 web dashboards).
Stack: Flutter, Firebase (Functions, Firestore, Auth), Claude API
Key challenges:
- AI assistant integration with healthcare-specific prompt engineering
- Sub-300ms response times for AI chat features
- Cost control at $0.02/user/month for AI inference
- Firebase Cloud Functions as AI gateway with conversation history in Firestore
- 4 separate Flutter apps sharing a common codebase through packages
Architecture decisions:
- Firebase Cloud Functions as a thin gateway to Claude API rather than direct client-to-API calls (security, cost control, rate limiting)
- Conversation history stored in Firestore with TTL-based cleanup to manage storage costs
- BLoC pattern for state management across all Flutter apps
Project 2: Web3 Token Launchpad (Solana)
A token launchpad platform with AMM (Automated Market Maker) using Meteora Dynamic Bonding Curves, real-time blockchain indexing, and AI-powered recommendations. Three separate codebases: Next.js frontend (1,454+ TypeScript files), NestJS backend, and FastAPI Python backend.
Stack: Next.js 14, NestJS 11, FastAPI, Solana, PostgreSQL, MongoDB, Redis, Qdrant, AWS ECS
Key challenges:
- Real-time blockchain data synchronization — indexer fetches pools, token metadata from Arweave, and prices from CoinGecko every 60 seconds, with WebSocket subscriptions for live swap detection
- RPC cost explosion: 26,000+ daily RPC calls to Solana (getAccountInfo), consuming 32.6% of QuickNode Build tier
- Bot traffic: 14,000 calls during 1-5am window (48% of daily total) from automated scrapers
- Dual-backend coordination: NestJS for blockchain indexing and pool API, Python for auth, AI agents, and token economy
RPC optimization (the biggest technical win):
- Redis balance proxy with 60s TTL — 90% reduction in getAccountInfo calls
- 5-tier bot detection middleware with behavioral analysis — 96% of bot traffic blocked
- Rate limiting (40/min, 500/hour) with graceful degradation
- Result: 26,000 → 2,800-5,000 daily calls (80-90% reduction), saving ~21.6M monthly credits
Architecture decisions:
- Contract-first API design: OpenAPI specs as single source of truth, frontend types auto-generated from backend
- Dual database strategy: MongoDB for flexible document storage (user profiles, tasks), PostgreSQL for financial data requiring ACID guarantees
- Qdrant for AI-powered task recommendations rather than collaborative filtering (better cold-start performance)
- Graceful Redis degradation: app continues functioning even when Redis is down
- Multi-wallet support: Phantom, Solflare, mobile adapters
Monitoring: Sentry integration across all services, Prometheus metrics in NestJS, Winston structured logging, health check endpoints.
Project 3: Content Platform
AI content generation and optimization platform (this evolved into the Crest project documented separately).
Stack: Python, FastAPI, Next.js, PostgreSQL, Redis, Celery
Key challenges:
- Multi-model LLM routing for cost optimization
- Content pipeline automation with quality gates
- A/B testing framework for content variants
- 8-queue Celery system for task prioritization
Shared Engineering Practices
Several engineering patterns proved valuable across all domains:
Standardized monitoring: Health check endpoints, error tracking (Sentry), and deployment notifications were consistent across all projects regardless of stack.
CI/CD templates: GitHub Actions workflows were templated and reused. A Flutter deployment pipeline written for the healthcare project was adapted for any future Flutter projects in under 30 minutes.
Cost tracking: Every project had explicit cost budgets for cloud resources and AI inference. Weekly cost reviews caught overruns early — a MongoDB index issue in the Web3 project was causing $40/month in unnecessary compute.
Documentation standards: Each project maintained CLAUDE.md files documenting architecture decisions, code patterns, and forbidden zones. This proved valuable when revisiting decisions months later.
Results
- 10k+ active users across healthcare and Web3 platforms
- 80-90% RPC cost reduction on Web3 platform through caching and bot detection
- 9+ repositories managed with consistent CI/CD, monitoring, and documentation patterns
- Pattern reuse across projects reduced setup time by roughly 40%
- Zero critical production incidents during the delivery period
Key Learnings
Patterns compound. The monitoring setup from the healthcare project took 2 days. Adapting it for the Web3 project took 3 hours. By the content platform, it was a 30-minute copy-and-configure. Investing in reusable patterns pays off quickly across projects.
Cost awareness prevents surprises. Weekly cost reviews across all projects caught 4 separate cost overruns before they became significant. The RPC optimization alone saved thousands in monthly credits.
Domain diversity strengthens architecture instincts. Working across healthcare (compliance-sensitive), Web3 (real-time, financial), and AI (cost-sensitive, latency-critical) forced clear thinking about trade-offs. Each domain taught patterns that improved the others.