Open-sourced July 2026 under Apache-2.0. MCP + REST, 68 tools, self-hostable or cloud-deployed — clone it and it boots with zero external services: no auth, no account, no API key. The system between foundation models and high-stakes decisions, under real constraints. Previously: reinforcement learning agents for autonomous discovery in manufacturing. The domain changed; the systems thinking didn't.
ase ↔ alchemi, crest ↔ alchemi; same tool call, open reference path or GPU path chosen per call), plus a pluggable auth / metering / audit spine so the same code runs standalone or hostedapi.novomcp.com for the OSS launchpip install novomcp-lite): eight zero-config tools, library or MCP server; same code as the full engine, minus the rest of the platformThe interesting problem isn't getting a model to say something. It's building the system around the model so its output can be acted on: safely, correctly, repeatedly, with a record of what happened. That's the work.
NovoMCP is what I built. One computational chemistry engine, 68 in-silico tools across ADMET, docking, molecular dynamics, quantum chemistry, structure prediction, and compliance, that speaks both MCP (reachable from Claude Desktop, Cursor, Codex, Zed, and any MCP-compatible assistant) and REST (any pipeline that posts JSON). Two paths to run it: self-host it under Apache-2.0 on your own hardware, or run it as a per-org cloud deployment with the certified FAVES compliance API wrapped around it. Same engine either way.
The engine, never the platform (positioning call). Scale-to-zero by default (unit economics call). Human-in-the-loop via MCP elicitation, not autonomous-agent-on-someone-else's-molecules (trust call). Open-sourced as an engine, not as a hosted product (moat call — the moat is FAVES-cert + operated cloud, not the code). See the full product portfolio →
Before NovoMCP: reinforcement learning agents for autonomous discovery in high-stakes manufacturing. The systems ran continuously, diagnosed their own failures, and improved without intervention. Different domain, same problem shape — the agent isn't the product; the system that makes the agent safe to deploy is the product. That's the throughline.
~/.novo/audit.jsonlA short selection of product, infra, and positioning calls made shipping NovoMCP. What I'd want a hiring manager to know is not what I built (that's the portfolio), but how I decide what to build, what to cut, and what to delay.
The straightforward call would have been to keep the engine closed and sell it as a hosted platform. I open-sourced it instead — Apache-2.0 top level, BSL 1.1 → Apache 2029-07-12 for the orchestration core. The moat is FAVES certification + operated cloud + trained models, not the code that calls RDKit and orchestrates docking.
Distribution flips from "sign up for our platform" to "clone the repo." Adoption compounds where the platform pitch couldn't reach — researchers, students, engineers wiring the engine into their own agents. The regulated buyers still need certified compliance; that's what they buy.
Every compute service ships with min_replicas = 0. Only the public-path surfaces (the gateway, the dashboard, the frontend) keep a warm replica. Pre-warm is an opt-in tier capability, not a default cost.
Carried the Azure cold-start posture forward to AWS rather than retrofitting it later. Unit economics stay honest from day one; the bill doesn't grow with idle.
"Platform," "OS," "infrastructure," "operating layer," "orchestration", every competitor in the category uses those nouns. They collide. I locked the public language to computational chemistry engine, with a diagnostic test on every public sentence: could a competitor put this on their homepage unchanged? If yes, replace the noun.
One word does the work of a positioning deck. Sales script, marketing copy, the assistant's own elicitation prompts, same word everywhere.
The original product had a campaign / iteration / quality-gate database schema for autonomous research runs. I cut it. Human-in-the-loop via MCP elicitation at every funnel stage replaced it. The user's AI assistant prompts the user at each gate.
Less code, clearer trust model, and the funnel runs inside the assistant the user already trusts, not as a standalone autonomous agent on someone else's compounds. The audit log is the quality gate now.
The prior model had a Free Trial / Core / Scale / Enterprise credit-tier ladder. I killed all of it at the OSS launch. Two paths now: Self-hosted, free forever (Apache-2.0, run it yourself) or Cloud deployment, contact for pricing (per-org subdomain, SAML SSO, GPU pool, certified FAVES). No self-serve middle tier.
The tier ladder was noise between the two real buyers — someone who wants to run the engine, and someone who wants us to operate it under an SLA with certified compliance. Removing the middle removes decision friction on both ends and cuts a bunch of billing complexity from the roadmap.
EKS API, Aurora, Redis, internal ALBs: all private. The catch: GitHub-hosted CI runners can't reach any of it, so every kubectl apply times out. Rather than poke holes, I committed to build on GH runners, deploy via OIDC → SSM → bastion and applied the pattern uniformly across every NovoServices repo.
Pattern doesn't re-litigate per port. Bastion comes pre-wired with kubectl, psql, EKS / Aurora / Secrets perms; new services pick it up for free. Security posture is a default, not an exception.
Research on how domain structure shapes reasoning - and how reasoning systems adapt under different objective and constraint regimes.
pip install novomcp-lite — eight zero-config tools (RDKit properties, profiling, synthetic-accessibility, PAINS/BRENK alerts, library screening, plus ChEMBL / ClinicalTrials.gov / bioRxiv search), as a Python library or an MCP server. Two dependencies, no keys. It's the exact code the full engine runs on, so the numbers match — just without the rest of the platform. Prompted by a researcher who wired the engine up to his agent and asked for a lighter build.Open to product, AI, and 0-to-1 leadership conversations. Bay Area or remote. Particularly interested in production AI systems, MCP-shaped infrastructure, and applied AI at companies where correctness matters.
Resume: in revision, available on request.