# 3 Days to a Production Support Chatbot A complete end-to-end run on k3ld Model β€” a support team fine-tunes a model on 12,847 helpdesk conversations, ships it as a live endpoint, and keeps it healthy. One person, ~4 hours of effort β€” at a fraction of what the same model costs the status-quo pipeline to produce. ## πŸ“… Day 1 β€” Connect + Explore (~15 minutes) - **09:00 πŸ”Œ Connect** β†’ add Managed S3 β†’ upload `support-assistant.jsonl` (12,847 conversations). Platform auto-detects schema. - **09:15 πŸ” Explore** β†’ search "support" β†’ table appears. Run `SELECT category, COUNT(*) … GROUP BY category` β†’ billing 4,201 Β· technical 5,892 Β· account 2,754. - **09:20 πŸ“Š Column profile** shows 3% null completions β†’ filter them out before training. Done β€” data is ready. ## πŸ“… Day 2 β€” Build: fine-tune + post-train (~1 hour, a fraction of status-quo cost) - **10:00 🧱 Fine-tune** β†’ base: Ministral-3-3B Β· data: `support-assistant.jsonl` (12,460 rows) Β· LoRA rank 16 / alpha 32 Β· ml.g5.2xlarge (spot). Click **Submit**. - **10:05** Status: submitted β†’ provisioning β†’ running. First metrics at 10:15: loss 3.42, throughput 1,247 tok/s. - **10:35 βœ… Training complete** β€” final loss 1.87. Test inference: "How do I reset my password?" β†’ "To reset your password, go to Settings β†’ Security β†’ Change Password…" - **11:00 🧬 Post-train** (optional preference tuning) β€” DPO on a few hundred hard negatives from the test set; ~30 minutes. - **11:30 πŸ“Š Eval** β€” built-in benchmark run ranks the checkpoint before shipping. ## πŸ“… Day 3 β€” Ship + Monitor (~2 minutes of interaction + one review) - **09:00 πŸš€ Ship** β€” one click from the completed run β†’ real-time endpoint. Platform packages model + tokenizer + inference script, creates autoscaling, and hands you an API key and a ready-to-call URL. Cost & capacity planner estimates instances and monthly spend before you deploy. - **09:15 πŸ“ˆ Monitor** β€” latency, drift, and usage dashboards are live. GPU auto-sleep stops billing when idle; the endpoint wakes on the first request. - **Weekly** β€” drift scan alerts on reply-quality degradation; retraining is one click away from the same dataset. ## Why it's fast The platform is the source of truth: catalog, training, checkpoints, deployments, and monitoring live in one system with one login and one audit trail. No Dockerfiles, no IAM debugging, no CloudWatch metric math, no "which checkpoint was best" archaeology.