Appendix - Capstones

What We Learned

• Integrate pretraining, post-training and evaluation into a reproducible model project.
• Serve under realistic load and defend optimizations using quality, latency and goodput measurements.
• Ship a tool-using agent with permission boundaries, approval, persistence and telemetry.
• Report baselines, uncertainty, costs, limitations and failure cases alongside the code.
• Explain architecture decisions and the business outcome with reviewable evidence.

Key Acronyms, Concepts & Jargon

Term	Short meaning
SFT / DPO / GRPO	Supervised Fine-Tuning / Direct Preference Optimization / Group Relative Policy Optimization.
MFU	Model FLOPs Utilization: useful model compute as a fraction of accelerator peak compute.
Model card	Documents intended use, training, evaluation and limitations of a model.
TTFT / TPOT	Time To First Token / Time Per Output Token: serving responsiveness metrics.
SLO / goodput	Service Level Objective / work completed within defined service targets.
Open-loop load test	Schedules arrivals independently of response completion to expose queueing.
MCP / HITL	Model Context Protocol / Human In The Loop: tool integration / human review boundary.
Durability / idempotency	Resume after interruption / retries preserve one intended effect.
Threat model	Maps attackers, assets, entry points and controls.
CI / pass^k	Confidence Interval / success across all k repeated trials.
ADR / ROI	Architecture Decision Record / Return On Investment: design rationale / business value measure.
Reproducibility	Enough versioned code, data, configuration and environment detail to repeat the work.

Back to section overview