i'm vega.

Affiliate-driven CRM for a Direct Selling Agent. Public capture forms and an EMI calculator feed leads into admin, department, and affiliate portals, each scoped by role — auth, RBAC, spam protection, transactional email, web push.

The interesting part is the constraint: it had to run at near-zero cost, so I built the async layer myself instead of paying for a queue. Background jobs drain off a Postgres table — each claimed so it runs once and retried if it fails — triggered by cron and Durable Object alarms. Turborepo monorepo, Next.js 16 on Cloudflare Workers via OpenNext, Neon Postgres with Drizzle. It survives a worker dying mid-job, and it scales to a real broker without a rewrite.

Built for the NIAT Agentathon. You type a query; a routing agent classifies intent and a dynamic orchestrator dispatches it across nine specialist agents — company scouts, risk hunters, analysts, debuggers, memo writers — assembling the answer as they go.

The engineering is in the failure handling. Llama 3.3 70B on Groq with a Gemini fallback, 429 backoff, and self-correcting retries that drop the temperature and re-ask when a model returns unparseable JSON. Every agent runs the same retry path, and the retries surface live — amber flashes on an animated agent graph.

FastAPI streams the whole run over SSE into a mission-control UI (React + Vite) wired to the real event stream, not a mockup. With no API key it falls back to deterministic mock data so the demo never dies on stage.

It started as an excuse to understand what actually happens underneath the async abstractions every web app leans on. I had wired up enough queues and cron jobs as black boxes; I wanted to build one from the metal up and find out what exactly-once, on-time, survives-a-crash really costs.

A workflow engine in Go: a pool of concurrent workers pulling tasks off a persistent queue, with retries and backoff, delayed and scheduled execution, and at-least-once delivery that survives a worker dying mid-task. The near-term goal is correct single-node semantics under load; the longer arc is distributed coordination — leases, heartbeats, and a leader election so the same job never runs twice on two machines.

It is early and exploratory, and I am building it in the open — design notes, the things that broke, and what I got wrong about concurrency the first time.