Short Answer

You want to add 5,000 Hindi medical terms to Llama 3 8B. Walk through the procedure and the risks.

A 1200-word system prompt, 200-word query, 150-word reply on GPT-5.5. Estimate the token cost, then name what doubles it.

You are building a 7B model for 25 languages including Hindi, Arabic, and Swahili. Design a fair tokenizer and name what it costs.

$50K compute budget, 50B tokens of US legal text, 1.5B parameter model. Defend your tokenizer design.

You are about to ship a feature that calls GPT-5.5 per request. Walk through how you estimate the token cost before launch.

Design an observability and debugging system for a production multi-agent application with 5+ collaborating agents.

Walk through SWE-bench, AgentBench, GAIA, and TAU-bench: what they measure, their shared blind spots, and how to interpret agent benchmark numbers.

In a few-shot prompt that extracts structured fields from invoices, why is a distinctive stop sequence (like a custom END_INVOICE marker) better than a paragraph break or the natural end of response?

Design an eval framework that tells you whether a production prompt survives a model migration (Claude 4.7 to 5.0, or Claude to GPT-5.5). What do you measure and what's the gate?

Explain how a DSPy-style 'prompt compiler' optimizer (like BootstrapFewShot) actually optimizes a prompt program and what makes it different from hand-tuned prompt engineering.

Design a production RAG system serving 10,000 concurrent users over a 100M-document corpus, with a P95 latency target of 2 seconds. Walk through the architecture and the bottleneck at each layer.

Describe a concrete, production runnable mechanism to detect hallucinations in a RAG answer (claims unsupported by the retrieved chunks), as the answer is generated or just after.

Design a golden eval set for a production RAG system serving a legal research product. Explain how you build it, how big it should be, what each example contains, and what you measure with it.

Your production RAG costs $1M/month. The CFO wants this cut in half with a max 1-point faithfulness regression. What highest leverage cost optimizations do you deploy, in priority order?

When IS fine-tuning the right call?

Three concrete situations where fine-tuning is the wrong tool

Non-obvious modes of train/test leakage in instruction-tuning evaluation

Tool-calling fine-tune: data shape and the common pitfall

Adding 500 domain tokens to the tokenizer: how do you initialise their embeddings?

Where does quality leak in a self-instruct synthetic data pipeline?

SimPO: what does it drop vs DPO, and what's the cost win?

Explain SFT's loss and the role of prompt-token masking

Sequence packing: what it does, and the attention-mask gotcha

Over-refusal: what it is and how naive safety FT causes it

Where do SFT and DPO sit in the classical RLHF pipeline?