Services architected
Reusable services I've architected and shipped across multi-tenant SaaS, fintech, AI platforms, streaming, and infrastructure. Each one is a real production workload, not a slide.
Three-tier Stripe Cashier integration with seat-based limits, upgrade and downgrade flows, and proration. Designed for clean tenant boundaries so a paying account never sees another tenant's invoices.
See it in CaseFlowStripe-style webhook delivery with HMAC-SHA256 signatures, retry on failure, signed timestamps to prevent replay attacks, and per-endpoint configuration. The pattern generalizes across every SaaS that needs outbound notifications.
See it in CaseFlowTenant isolation via a global scope on a provider_id column instead of subdomain isolation or a tenancy package. Lighter, easier to reason about, and survives audit because it lives in one place.
See it in CaseFlowMulti-step document capture with retry logic for fumbled photos, server-side validation, and an operator review queue. Fields and required documents are data-driven, not hardcoded.
See it in ZarpayData-driven anti-money-laundering rules that flag transfers based on amount, frequency, corridor, and recipient pattern. Operators review hits in a dedicated queue with audit log entries.
See it in ZarpayPluggable interfaces for payment-in, payout, FX, and OTP vendors. Going live behind a licensed counterparty becomes a provider swap, not a rewrite.
See it in ZarpayDocument upload, chunking, embedding, and hybrid retrieval (full-text + vector) over pgvector with tenant scoping. Hybrid search meaningfully outperforms pure vector at this scale.
See it in AxonModel Context Protocol servers (postgres + custom template) bridged into a LangGraph agent. Gives agents structured access to databases and internal APIs without leaking credentials.
See it in AxonStreaming chat with explicit tool schemas, retry handling on tool failures, and run logs that survive a server restart. More reliable than prompt-chained pipelines.
See it in AxonLong-running agent work runs in BullMQ workers, not in the request path. Includes a Bull Board admin UI for queue visibility and dead-letter handling.
See it in AxonTenant isolation enforced at the database role level via a dedicated app role and a withOrg pattern. Survives a leaky query because the database itself refuses to return cross-tenant data.
See it in AxonPrometheus metrics, Grafana dashboards, Loki log aggregation, and Langfuse for LLM-specific telemetry. Pays for itself the first time an agent burns tokens on the wrong path.
See it in AxonMulti-resolution HLS playback (480p, 720p, 1080p) with resume, variable speed, and graceful bitrate switching on connection changes.
See it in ReelmWorker pipeline that transcodes single-source uploads into the full HLS variant ladder, with live progress reporting back to the admin console.
See it in ReelmInline AI suggestions for priority, labels, and assignee on new issues, based on what the project has resolved before. The user confirms or overrides; the AI never auto-closes.
See it in TracklaneNatural language questions answered across every issue, comment, and wiki page in a project. Each answer cites the source thread so users can verify.
See it in TracklaneProduction-grade workflows with structured Claude tool-use, shared workflow_runs table for unified observability, and a single dashboard across multiple workflows.
See it in n8n Agent StudioTwo-sided bid lifecycle: rider posts a max budget, drivers submit bids, accept-the-best with rating context. Typed API client shared across both React Native apps.
See it in DriveBidSub-second search at scale with LLM-based reranking on top of Elasticsearch. Recall and relevance are split: the index does keyword and BM25, the AI layer reorders the top-K based on user intent. Holds up at multi-million-document indexes.
Multi-strategy data extraction with fallback selectors, deduplication, and adaptive throttling. Built to survive long-tail variation across upstream HTML structures and to back off cleanly under rate-limit pressure.
Email, SMS, and push delivery from a single fan-out service with per-channel rate limiting, retry, and delivery audit. One outbound contract across providers, swappable behind an interface.
Claude-driven optimization passes for content where hand-rolled rules stop scaling: keyword extraction, title rewriting, taxonomy assignment. Loop converges on better outputs than static rule sets at production scale.
Validation pipeline with extensive coverage (400+ tests) for redirect chains, query parameter normalization, and edge cases that surface at scale. Designed to run in CI and as a runtime gate.
Open-source Lua plugin for Kong that adds Redis-backed response caching with configurable TTL and shared cache across gateway nodes.
See it in Kong Redis PluginShopify Partner-listed embedded apps with OAuth scope tracking, HMAC-verified webhooks at the framework layer, and a Polaris-native admin so the app feels native to the Shopify shell, not bolted on.
See it in StoreBridgeTenant isolation enforced by Postgres itself via a non-privileged app role and RLS policies on every tenant-scoped table. Test suites prove cross-tenant reads and writes are rejected against a live Postgres, not a mock.
See it in StoreBridgePoints, badges, and leaderboards wired into core learner activity (course completion, quiz scores, streaks) with admin-configurable rules. RBAC for instructors, learners, and org admins.
See it in LearnLoopThree-tier subscription wiring with Stripe Checkout, the customer billing portal, and plan limits enforced at the model layer (not just the UI) so a free-tier user cannot exceed quotas via the API.
See it in LearnLoopOne repo that stands up a production-grade Odoo 18 instance on AWS, Azure, DigitalOcean, or local Docker. Terraform modules, hardened compose, demo data seeders, and a free-tier Oracle Cloud variant for a $0 demo URL.
See it in OdooForgeNative Odoo addon that adds a tool-using AI support agent inside the helpdesk module. pgvector retrieval, Claude or Groq provider abstraction, and a filterable audit log on every AI action so operators see exactly what the agent did.
See it in OdooForgeParallel dispatch to multiple upstream APIs with per-provider adapters, normalizers to a canonical schema, dedupe at the result store, and circuit breakers so one slow provider does not poison the others. Streams to the client over WebSocket so users see results as each provider responds.
See it in SkylaneDocument Q&A pipeline with chunking, OpenAI embeddings into Pinecone, and a LangChain retrieval chain that returns answers with source citations. Streaming UI so the answer renders token by token.
See it in ChatPDF