Building a DL Model in 3 Weeks with Minimal Coding

By leveraging AI agents, developers can build DL models for search without directly typing training code, evaluation code, or data generation scripts. The workflow involved reviewing commands generated by the agent, executing them manually, and feeding raw server logs back to the agent for root cause analysis and fix suggestions.

Server access was restricted to humans, while repetitive tasks were fixed in scripts. By separating sessions from training and applying checkpoints, memory watchdogs, and a 20-step rehearsal before full runs, 3-hour training cycles were reliably repeated on a single GPU server.

The core bottleneck for performance was not the model architecture but the data. Contradictions where the same item appeared as both correct and incorrect answers were removed, and missing search types in actual logs were supplemented with LLM synthetic data.

• Type definition → Mass generation → Review → Filtering & regeneration
• During training, papers were read, and loss functions and formulas were supplemented via the agent
• Tools like Paper2Code and academic-research-skills were used to automate paper research and implementation

The developer's role was to define what to build, question anomalous results, and restrict code changes that could compromise reproducibility. When introducing new models or tools, regression tests with fixed baseline metrics were run first to distinguish the impact of experimental changes from tool changes.

Markdown documentation was automated to be uploaded directly to the wiki by the agent. The key takeaway is not merely writing less code, but rapidly iterating on the experiment loop—from environment setup to verification and documentation—and filling knowledge gaps with AI.