AI Projects › Six AI skill areas
Each area states what was built, the lesson it taught and an honest status. Where something is designed but not running, this page says so.
| Area | Status (Nov 10) |
|---|---|
| Guardrails | Proven in production |
| Agents | Running |
| LLMOps and operations | Running |
| Evals and quality gates | Proven, extending |
| RAG and retrieval | Designed, groundwork done |
| Fine-tuning | Deliberately deferred |
Guardrails
A five-level risk ladder, verbatim approvals, read-only agents, a STOP word, dry runs before every go and identity-gated destructive commands. See guardrails. Lesson: a rule written the night the mistake happened beats a rule written in a planning meeting.
Agents
A manager agent that plans, verifies and documents, plus two worker agents in containers. Job cards list inputs, process and outputs, and a 24/7 dispatcher runs check-backs. Lesson: agents hold stale briefs, so durable cards beat relaying messages by hand. Still open: a message bus for agent-to-agent handoffs is designed, not built.
LLMOps and operations
A run console for every long job, collectors that need no install on the machines they watch, and orchestration in plain scripts that use zero model tokens. Local GPU models are planned for bulk work, with Claude managing and sampling the output. Lesson: cost is a design input. One AI plan was shared by three agents.
Evals and quality gates
VMAF gates on every encode, per-volume scan gates (G1 to G3), and warm-up laps before promotion: 32 of 32 files independently re-hashed with sha256, size and modification time matching after six defects were fixed. Reporting always shows pass-only savings and whole-pool savings together. Lesson: the gate measured pixels and missed metadata, and prompt-tuning a local model made its wrong answers more confident, not more correct.
RAG and retrieval
Groundwork done: a fingerprint for every file, a media probe for 683,000 files (codec, resolution, GPS, camera make) and full lineage, with Postgres and pgvector in place. Design: hybrid search over filename and path, transcript text and vector similarity, with every result keyed by sha256 so a moved file never breaks a link. Status: designed, not yet in production. Lesson: fingerprint first, then index.
Fine-tuning
No fine-tuning until evals exist and the dataset is clean and de-duplicated. Training on twenty duplicate generations of the same clip would teach the wrong thing. A clean, labelled corpus (by project, client type, camera, location and date) is a by-product of this work. Lesson: knowing when not to train is part of the skill.
Also proven, outside the six
Content-addressed inventory, data lineage, golden-record selection, information lifecycle and tiered storage.
Industry terms, mapped to proof
| Term | Where it is proven here |
|---|---|
| Agentic workflows, multi-agent orchestration | Manager plus two workers, cards, dispatcher, the walk-away morning |
| Human-in-the-loop, approval gates | The risk ladder, with levels 3 and 4 human-run and 113 verbatim approvals |
| Guardrails, least privilege | Read-only agents, denials as hard rails, identity-gated destructive commands |
| LLMOps, observability | Run console, collectors, ETAs, status colors, check-ins |
| Evals, quality gates | VMAF gates, G1 to G3 scan gates, warm-up laps |
| RAG, hybrid search, vector database | Designed on pgvector, groundwork done |
| Context engineering | Layered context, stage folders, one home per fact |
| Data lineage, provenance | Parent-to-child sha256 links for 3,501 encodes, 98.9% traced |
| Master data management, golden record | Health and completeness first, recency as a tiebreaker |
| Data quality, deduplication | sha256 identity for 14.7 M files, copy map, truncation detection |
| FinOps, information lifecycle | Protect, reclaim, compress, archive, with reclaim in TB and dollars |
| Root cause analysis, postmortems | Boot-panic diagnosis and the 360 root cause |
Statuses are as of Nov 10 and follow the project’s own labels. Nothing is described as in production unless marked running.