Explore distributed LLM inference codebase (HyperCluster)
Takeaway: I'd start with the same broad 'explore everything and list features' prompt, since this was a pure read-only investigation and it worked fine end to end.
Stack · 19 sessions
transformers sessions with coding agents
19 sessions on transformers with a coding agent: the agent failed at least once in 13, and 6 finished without a human stepping in.
Fine-tune RoBERTa on biglam/on_the_books via HF Jobs
Takeaway: Next time I'd pin the transformers/trainer API version up front, since most of my failed jobs were just Trainer argument mismatches (evaluation_strategy, tokenizer vs processing_class) burning GPU minutes on HF Jobs.
Deploy TutorDesk AI on Modal + Gradio with Qwen3-4B and MiniCPM-V
Takeaway: Next time I'd tell it upfront which library versions are pinned (transformers, TRL) since half the debugging loop was chasing breaking API changes across versions I hadn't pinned.
Write model card for phonetic ASR evaluator model
Takeaway: Next time I'd tell it upfront that the eval reuses training examples so it doesn't have to discover that itself, and that .gitignore excludes models/ so it can plan the tracked-vs-mirrored file split from the start.
Build Modal ASR/TTS backend for a Gradio read-along app
Takeaway: Next time I'd give the exact test fixture (silent WAV) up front so the agent designs the endpoint to handle it from the start instead of hitting a gated-model load in production.
Fix over-prescribing Action Plan prompt in student risk system
Takeaway: Next time I'd ask upfront for the prompt's exact output-format constraints, not just the risk-level instructions, since the rigid 4-week structure was quietly forcing intervention-heavy plans regardless of risk.
Add GraniteMoeHybrid architecture to mlx-lm
Takeaway: Next time I'd front-load the transformers layer forward-pass source (not just the module files) so the agent sees the double-residual MoE/shared_mlp structure per layer up front instead of discovering it through hours of logit-diffing.
Deploy and debug Modal endpoints for a Hugging Face TUI Space
Takeaway: Next time I'd start the session by asking the agent to check both Modal deployments end-to-end (health + authenticated /chat) before I go poking at env var naming one variable at a time, since most of the friction was just me and the agent working through STORY_FORGE_API_KEY vs MODAL_API_KEY confusion.
Survey ML datasets, models, and deployment options via Q&A
Takeaway: This wasn't really a build — I used the assistant purely as a research reference to survey datasets, models, fine-tuning approaches, and deployment options before deciding on my own stack.
Survey ML models, datasets, and infra options for an ML devtools dashboard
Takeaway: I'd next time scope this to one concrete task with the actual repo attached, because the agent kept hitting sandbox restrictions and defaulting to generic textbook answers instead of anything grounded in our code.
Fine-tune ModernBERT on Jim Crow law dataset via HF Jobs
Takeaway: Next time I'd mention up front that the transformers API had renamed tokenizer= to processing_class=, since that's what caused the one failed job run.
Survey LLM/ML tooling, datasets, and deployment options via open Q&A
Takeaway: I'd start by scoping this to the actual decisions I needed to make for my project instead of asking the assistant to cover the entire ML ecosystem in one sitting — most of these answers never got applied to code.
Survey datasets, models, and deployment options for an ML/LLM stack
Takeaway: I used this as a rapid-fire research session rather than a coding task, so there was nothing to intervene on or fix — it was just me pulling information across a huge surface area.
Survey ML tooling: datasets, models, deployment, fine-tuning
Takeaway: I'd frame this upfront as a single research brief instead of firing off 250+ standalone questions, since the assistant kept resetting context and repeating boilerplate advice rather than building on prior answers.
Survey ML tooling, models, and infra options for an AI dashboard repo
Takeaway: I'd start by telling the agent up front which of these ~130 questions actually matters for the repo I'm working in, instead of firing off a long, unrelated stream of general ML trivia.
Survey ML tooling and model choices via Q&A chatbot
Takeaway: I'd start by giving the agent actual repo access or pasting in the relevant files myself, since half these turns were it telling me it couldn't read anything due to sandbox restrictions.
Debug Modal QLoRA fine-tuning of MiniCPM-V-4.6
Takeaway: Next time I'd start by checking the model's actual transformers/peft requirements against unsloth's supported range before writing any training code — that single incompatibility cost the whole first hour.
Push fine-tuned MiniCPM-V invoice model to HuggingFace and fix inference format
Takeaway: Next time I'd start by checking the app's actual inference stack (llama-cpp-python, not vLLM) before pushing any model format, since that assumption cost a lot of back-and-forth on GGUF conversion that turned out to be a dead end for MiniCPM-V-4.6.
Fix MiniCPM-V fine-tuning pipeline on Modal with Unsloth
Takeaway: Next time I'd tell the agent upfront to check the model card on Hugging Face for the exact transformers version needed, since half the debugging loop was chasing version mismatches (MiniCPM-V-4.6 model_type not registered, then peft/transformers pin conflicts) that a single doc check would have caught.