⚡ What I do

Software is going through its biggest shift since the internet itself — the next generation of products won't be built by large teams writing boilerplate, they'll be built by engineers who can fuse AI, systems design, and product instinct into one skillset. That's the bet I'm making.

I build full-stack products where the backend is an AI system, not a feature bolted onto one. That means treating LLM calls like any other unreliable network dependency — timeouts, retries, circuit breaking, structured logging, graceful degradation — while still shipping a UI people actually want to use.

scrape → LLM analysis → resilience layer → score → stream → ship

Right now: agentic data pipelines on serverless infra, production reliability for LLM-backed APIs, and the unglamorous plumbing that keeps AI features from falling over under real traffic.

Focus areas: Multi-Agent Architectures Agentic Workflows RAG Systems Knowledge Graphs LLM Applications AI Infrastructure System Design Distributed Systems Backend Architecture Developer Platforms Platform Strategy Product Engineering

🏰 DevCastle — Featured Build

A developer platform combining AI-native workflows, market intelligence, and builder infrastructure. Live. Deployed. Hardened against real failure, not a demo.

📊 Reddit Market Intelligence Agentic pipeline: scrapes target subreddits → LLM-driven analysis → ranked SaaS opportunity scoring. Built to survive serverless constraints and flaky upstream APIs. SSE streaming — long AI jobs stream incrementally to dodge Vercel's 504 ceiling InsForge AI resilience layer — timeouts, backoff retries, error classification, auto credit-refunds on failed-but-charged requests Aiven MySQL retry logic — handles transient connection drops from cold-start reconnects Structured JSON logging → Vercel Log Drain / Axiom, full request traceability Test coverage — Vitest + curl smoke tests against the live API

🔌 G2 API V2 Integration Full market-intelligence integration: buyer intent, competitor analysis, opportunity scoring, category browsing, credit management. ~22 files · ~2,400 LOC of TypeScript Diagnosed a production 401 traced to a missing env token Fixed silent auth failure: Token token= → Bearer — the kind of detail that only surfaces by inspecting raw traffic, not docs Built an in-app API explorer so credit-metered calls can be tested without burning quota blind

Stack: Next.js 14 (App Router) TypeScript Tailwind CSS Prisma Aiven MySQL InsForge AI G2 API V2 EmailJS

Platform pillars: AI-native workflows · intelligence layer (startup discovery, competitive analysis, trend detection) · builder infrastructure (developer communities, project discovery) · platform engineering (scalable, cloud-native architecture)

🔭 Other builds

🤖 Multi-Agent AI Assistant Production-grade AI orchestration powered by LangGraph — agents that plan, delegate, and execute as a coordinated system rather than a single prompt-response loop.

🔍 AI Knowledge Systems Retrieval-first architectures for intelligent information discovery — RAG pipelines built for accuracy over vector-search vibes.

🧠 How I debug

I validate API behavior directly against the live service — curl, raw request/response inspection — before trusting an SDK's abstraction or assuming the docs are right. The hard bugs in AI-integrated systems live in the gap between what a provider claims and what it does under load, rate limits, or cold starts.

## 🛠️ Tech Stack

_{Core stack used in production (Next.js · TypeScript · Prisma · Aiven MySQL · InsForge AI) marked in the DevCastle section above — the full list below spans what I build with day-to-day across full-stack and AI-native engineering.}

Frontend

Backend & APIs

Data & Storage

AI / Agentic Engineering

RAG pipelines Multi-agent orchestration Agent memory & state Tool/function calling Prompt engineering Embeddings & semantic search LLM evaluation & guardrails Streaming inference (SSE)

Infra, DevOps & Observability

Serverless architecture Structured logging Retry & circuit-breaker patterns Rate limiting Cold-start optimization Log aggregation (Log Drain)

Tooling & Workflow

Integrations shipped in production
InsForge AI G2 API V2 EmailJS Vercel Log Drain OAuth providers

📐 Engineering principles

Principle	What it means in practice
LLM calls are unreliable I/O	Timeouts, retries, fallbacks aren't polish — they're what separates a demo from a product
Observability before scale	If a prod failure can't be traced request → root cause, it isn't done
Verify, don't assume	Test against the real API/DB/service before believing a fix works

📬 Let's talk

Open to conversations on agentic system design, production AI reliability, and full-stack architecture.

📊 GitHub Stats