Smarsh
Forward Deployed Agentic Engineer
Go, HTMX, Rust (desktop), Shell, Postgres, pgvector, Docker, vLLM, haproxy, nginx
- Stood up SmartAI - the OS company agents live in. Shared memory, versioned playbook, gated tools, named goals. One Go binary. Dark factory runs the work.
- Control Tower is the god's-eye view: live runtime visibility for high-level decisions.
- Turned Snowflake into a live warehouse Goal: scheduled ingest into PostgreSQL, delta updates, intelligent content merge — not a one-off dump.
- Post-trained smaller self-hosted LLMs with human-in-the-middle RL feedback; so they become experts at the organization's domain work. Eat the token cost once and own the models.
- Added intelligent routing so requests hit the right model based on what the user actually needs.
- Used customer data signals coming through the system to surface churn risk earlier.
- Custom-built a graph of the team brain with a realtime heat map that holds millions of nodes without falling over.
- Stacked it on Postgres + pgvector for retrieval, with Go services and an HTMX UI. Rust for the desktop client.
Key achievement. SmartAI. One Go binary. Agents live in it. Dark factory runs the Goals. Control tower is the god's-eye. Smaller self-hosted models get better at the company's work because humans sit in the loop. First use case: spot at-risk accounts to protect margin and cut churn. Dark factory is how the next use case ships without standing up a new system.