A fine-tuned Qwen3-8B agent with a Fast/Slow harness and Temporal-backed durable workflows that act for the user across phone, browser, and email.
Concept artwork
An LLM post-training and durable agent platform: Qwen3-8B fine-tuned via SFT + 4-bit QLoRA on provenance-tracked agent trajectories, a Fast/Slow harness that separates low-latency voice interaction from long-horizon planning and tool use, and Temporal + PostgreSQL workflows that survive worker restarts.
Achievements
Fine-tuned Qwen3-8B via SFT + 4-bit QLoRA on 1.5K provenance-tracked agent trajectories synthesized and filtered by a Claude Sonnet teacher; lifted held-out task completion 58%→67% and cut false completion 6%→2%.
Architected a Fast/Slow agent harness separating low-latency voice interaction from long-horizon planning and tool use, with shared state, retries, approval gates, and completion checks across phone, browser, and email.
Implemented durable long-running agent workflows with Temporal + PostgreSQL, persisting approval waits and scheduled follow-ups while using idempotent retries and callback deduplication to recover across worker restarts.