Drawn from the actual repos (NewsRx/hermes, hermes-sme-fe, optinosis) — file-evidenced, not resume claims · Answers: UCB_Interview_Answers.html
| Category | Now — NewsRx + Optinosis (2024–2026) | Past — career (2002–2025) |
|---|---|---|
| Agentic coding | Claude Code (primary), Kimi Code, OpenCode agents (OptiBot/BirdDog) wired to local vLLM endpoints; Claude Code + Context7 + Playwright MCP servers | Pre-agentic era — Gracie chatbot (Dialogflow, 2017); everything else traditional SDLC |
| Models / inference | Qwen3-32B on vLLM/RunPod (production); Qwen3.6-35B FP8, 262K context on RTX PRO 6000 Blackwell (Optinosis); Gemma 4 31B MLX 4-bit on-device; Claude Sonnet 4.5 as LLM judge; Ollama (deprecated 2026) | Supervised ML, NLP, predictive analytics (Invistics 96% accuracy); SageMaker-era cloud ML |
| Eval / LLMOps | HHEM-2.1 faithfulness, BERTScore (DeBERTa), readability; blinded 4-clinician persona QA harness w/ inter-rater agreement; MLflow tracking + registry; experiments persisted to Postgres | Model validation under Lean Six Sigma; Monte Carlo; no LLM evals (pre-LLM) |
| Language / frameworks | Python 3.14 (uv, hatchling, Pydantic v2), FastAPI, SQLAlchemy 2.0, Polars/pandas; React 18 + Vite + MUI (Optinosis front ends) | Python, SQL, R; SAP/JDE ecosystems; FORTRAN (early career) |
| Data layer | AWS RDS PostgreSQL — sole DB, dev/prod schemas, Alembic sole DDL manager; bronze/silver/gold medallion; JuiceFS shared FS; MinIO PHI store + CNPG Postgres on K8s (Optinosis); DVC on DO Spaces | Oracle, Cloudera→AWS Redshift migration (J&J), Snowflake (Optum), Databricks + Alation (Evernorth), SAP/JDE feeds |
| Frontend / apps | Streamlit SME review instrument (auth, audit log, schema isolation); React dashboards (OptiStrata/Opti-Anthropos) | Tableau, Power BI; Salesforce Visualforce/Apex (Street Grace) |
| Infra / hosting | Render (2 services), RunPod GPU; self-hosted K8s — Talos + ArgoCD GitOps, Kubeflow Pipelines, KServe, Cloudflare/Tailscale | AWS (S3, Redshift), Oracle Cloud, on-prem enterprise (Cardinal, 80+ sites) |
| Testing / CI | pytest, Playwright QA harness; GH Actions gates — doc-lint, DDL-parity, required review; pre-commit hooks; Stryker mutation testing; schema-isolation integration proofs | Enterprise UAT ownership, SAFe test governance; no modern CI (pre-era) |
| Governance | CLAUDE.md / AGENTS.md agent instructions under lint CI; SOC 2-aligned controls; fail-closed auth; audit-log tables; PHI-perimeter design; Sealed Secrets/Vault | GxP-validated systems (Amgen first-of-kind GenAI), change control, Alation semantic layer, FDA/NRC audit trails |
~60-second answer: "Python, and I build with agents — Claude Code leading, Kimi Code and OpenCode in the rotation, wired to local model endpoints. Production inference is Qwen3-32B on vLLM over RunPod; on the bench I've run Gemma 4 31B on MLX and Qwen 3.6 at 262K context on a Blackwell workstation. Evaluation is bespoke and serious — HHEM faithfulness, BERTScore, a blinded four-clinician persona harness, everything tracked in MLflow. Data layer is AWS RDS PostgreSQL, Alembic as the only DDL path, medallion architecture, dev/prod schema split enforced by tests. Apps ship on Streamlit and Render. Underneath the code: GitHub Actions gates — documentation lint, DDL parity, required review — because agent-written code gets the same governance as human-written code. Before this it was two decades of Oracle, Redshift, Snowflake, Databricks, SAP — model-agnostic, cloud-agnostic, everything under version control and test."