SKILLS · 36 · EVIDENCE-LINKED
Nothing here that I can't open the file for.
Each skill links to the projects that use it. Counts come from the case studies, not from me.
Lime = core· ×N = projects, open to list
Orchestration & agent control
Seven skills across five projects. Five are core.
Multi-agent orchestration (custom)×2
Claude Agent SDK & subagents×3
Prompt engineering, agent briefs as versioned code×4
Human-in-the-loop approval gates×2
Compliance-as-code×1
Adversarial verification panels×1
LLM cost & quota engineering×1
Retrieval & grounding
Seven skills across five projects. One is core.
Structured / JSON-schema output×2
Google Gemini API×3
Pydantic×2
ONNX Runtime (MiniLM embeddings)×2
LangChain (document loaders, text splitters)×1
Gates, provenance & audit
Six skills across four projects. None are core.
Application security, JWT, bcrypt, TOTP MFA, Fernet encryption, RBAC×2
Semgrep & Bandit integration×1
Build manifests, data contracts & runtime assertions×1
Append-only audit & approval ledgers×3
pytest, 215 tests, CI-gated×3
GitHub Actions CI/CD×1
Data & pipelines
Eleven skills across eight projects. One is core.
Data pipelines / ETL×2
PostgreSQL / Supabase×2
SQLAlchemy 2.0 + Alembic×1
scikit-learn / LightGBM×1
Model evaluation (class imbalance, F1/precision/recall)×2
- Power BI×1
- Transformer fine-tuning (token classification)×1
- Hugging Face Transformers×1
- NER / token classification×1
Interfaces & delivery
Five skills across four projects. One is core.
FastAPI×2
React 18×2
TypeScript×2
Docker×1
Cloud deployment (Render · Vercel · Supabase · Cloudflare R2)×1
NOT YET
No LLM-output evaluation, no GPU serving, no managed vector index at scale, no distributed tracing. The nearest one is measuring my own verifier. The harness is designed, answer keys anchored to verbatim source strings so a hallucinated location dies to a failed string search, a human adjudicator so an LLM never grades its own output, a held-out split touched once, and building it is what I'm doing next.