I ship frontier AI from prototype to production: agents, retrieval (RAG), evaluation, and the integrations that make them stick.
Building agent reliability tooling and post-RAG retrieval at scale; selectively open to fractional advisory through the next quarter.
Each row links to a public repository and cites a concrete outcome.
| Project | Demonstrates | Stack | Status |
|---|---|---|---|
| promptsheon | Git-native, versioned infra for agent configs — reproducibility and governance across 5 internal teams. | TypeScript | active |
| agent-passport | Identity and provenance for production AI agents — adopted by 2 enterprise rollouts in 2025. | Python | active |
| agent-guard | Reliability and safety controls for deployed agents — reduced runaway-action incidents by ~70% in pilots. | Rust | active |
| delta-search | Retrieval over changing corpora — operates a 50k QPS index with p99 < 800ms. | Go | active |
| underwrite | Decision systems on production data pipelines — turns 3-week eval cycles into < 1 day. | Python | active |
| fleetpilot | Distributed event-driven backends — resilience and scale patterns for cloud-native platforms. | Go | maintained |
Browse the full public repositories list.
- Eval harnesses for LLM systems — frameworks used across three production agent rollouts; cut eval-suite iteration time from days to minutes.
- Production RAG — retrieval, re-ranking, and freshness pipelines at 50k QPS with p99 under 800ms (2025).
- Agent reliability — identity, guardrails, and provenance for long-running agent workflows.
- Applied architecture — translating business outcomes into tractable AI system designs.
Daily: Python, TypeScript, FastAPI, Postgres, OpenAI API, evals tooling. Weekly: Rust, Go, Kubernetes, retrieval infra, OpenTelemetry. Occasionally: PyTorch, vLLM, Hugging Face, CUDA.
- Architecture & advisory — aligning AI investments to business outcomes across enterprise and digital-native customers.
- Engineering & implementation — hands-on systems work: agents, retrieval, embedding pipelines, APIs, SDKs.
- Security & governance — designing and deploying within enterprise security, privacy, and data-governance requirements.
- Ecosystem enablement — reference architectures, workshops, and reusable patterns for partner and internal teams.
Most of my work is public. I default to building repeatable patterns rather than one-off solutions.
- Open to: fractional advisory, paid pilots (4–8 weeks), and select full-time consulting engagements.
- Not open to: cold 0→1 MVPs without budget, "let's chat" intros without context.
- Response time: typically 2–3 working days.
- Best email format: one paragraph of context, the question, and a proposed next step.
- ✉️ Email: sachncs@gmail.com
- 💼 LinkedIn: linkedin.com/in/sachncs
- 🐙 GitHub: @sachncs
- 📅 Advisory calls: request via email with a 30-min agenda
If what you've read resonates, reach out.