Wycliffe Rotich

Wycliffe Rotich

AI infra's part nobody wants to build
#101549950 followers0 following
>10,000All time
6 KP

About

Backend and distributed systems engineer based in Nairobi, 7+ years across enterprise systems, cloud platforms, and AI infrastructure. I build things that stay correct under real failure conditions, not just in the happy path. Currently building AetherGrid, an open-source distributed AI workload orchestrator with lease-based execution ownership, real subprocess execution, and reconciliation that actually recovers from crashed workers and dead nodes. I also evaluate LLM-generated code for correctness and architecture as an AI training specialist. Always up for talking distributed systems, clean architecture, or anything that breaks under concurrency.

Badges

Tastemaker
Tastemaker

Forums

AetherGridp/aethergrid•

18h ago

Building AetherGrid: an AI orchestrator that survives real worker failures

I've spent years chasing race conditions in production systems, revenue systems handling real money, Kafka worker queues, and now reviewing LLM-generated backend code for a living. The pattern is always the same: code works fine until something fails mid-execution, then two processes think they own the same job.

AetherGrid is my answer to that. Job execution ownership runs on time-bound leases that renew continuously for the job's whole runtime, not a boolean flag that lies to you the moment a worker crashes or a heartbeat drops. Real subprocess execution with enforced timeouts, reconciliation that reclaims abandoned work, 280 tests behind it, including one that genuinely kills a process that ignores SIGTERM.

Live now on Product Hunt, launching August 21st. Would love feedback from anyone who's dealt with execution ownership or lease-based systems in their own infra, and happy to answer anything about the architecture.

Repo: https://github.com/wycliffRotich...

•

3h ago

AetherGrid - Open-source AI workload orchestrator, lease-based execution

AetherGrid is an open-source AI workload orchestrator built to survive real failure, not demo well. Jobs run under time-bound, continuously-renewed leases, not a simple assignment flag, so a crashed worker can't cause duplicate execution. Real subprocess execution with enforced timeouts, reconciliation that reclaims abandoned work, and 280 tests proving it, including one that force-kills a process ignoring SIGTERM. Clean Architecture, Domain-Driven Design,26 ADRs, live with CI/CD on every merge.
View more