AI Product Manager

Responsibilities

• Platform Strategy and Problem Selection: Owning a platform-first applied AI roadmap: choosing which problems deserve AI at all, shaping reusable primitives — knowledge retrieval, agent tooling, gateways, access infrastructure — that multiple products and teams compound on, and recommending conventional software where that is the better answer.

• Hands-On Product Development: Developing AI products by writing the production Python behind them rather than prototypes alone, while collaborating closely with DevOps, SRE, security, and product teams so what ships is operable, not just demonstrable.

• Verification and Quality Bars: Testing capabilities hands-on before they ship or get bought: reconciling analytics against the system of record, running build-versus-buy analysis on measured economics, and setting the bar that AI proofs-of-concept must clear — action boundaries, data access, security — before anything reaches production.

• Stakeholder Collaboration and Adoption: Serving as the liaison between technical teams and business stakeholders, carrying initiatives from requirements through rollout, documentation, onboarding, and support — adoption is part of the product, not an afterthought.

• AI Governance and Spend: Bringing the AI estate under governance: per-project cost attribution, subscription consolidation, an intake path for new AI tools, and security escalations when a risk crosses team boundaries.

• Hiring and Team Building: Leading hiring for AI-engineering roles end to end — role definition, practical tasks, technical interviews — and holding the bar rather than hiring down.

Achievements

I operate this role with a deliberately platform-first posture: build reusable primitives — knowledge retrieval, agent tooling, gateways, access infrastructure — and let multiple products, teams, and workflows compound on top of them. Alongside the AI products delivered under this role (linked from the Related Projects panel), what follows is how those products get built and the cross-cutting outcomes that made them land at company scale:

• Hands-on delivery of the production AI stack. The retrieval backbone and the agent-facing MCP tooling are not only products I scoped — I write the bulk of their Python myself: an async FastAPI service, LangGraph retrieval pipelines over a vector store with provider-fallback embeddings and cross-encoder reranking (including an additional reranking stage that raised relevance and cut noise), FastMCP tool servers with per-request personal-access-token authentication, search and fetch tools that put the same retrieval in front of ChatGPT's knowledge and deep-research surfaces, the pytest suites that cover it all, the GitLab CI pipelines that run those tests before any tagged release deploys, and the Helm values that ship both services onto Kubernetes with horizontal pod autoscaling. LangSmith tracing is wired through the retrieval path, so request-level behavior is observable in production.

• An LLM gateway operated as a product. I built and operate the corporate gateway that sits between every engineer's coding agent and the upstream model providers — adapted from an open-source base and rebuilt into a corporate-grade service several times its original size. It does provider routing and failover across multiple AI labs, per-user request, token, and cost accounting priced at API rates — built so the economics of any usage model stay decidable — an authenticated admin surface with usage and cost analytics, a versioned client contract that tells an incompatible client what to do instead of generating support load, and centrally administered usage governance. It kept serving through a year of continuous upstream change — contracts, pricing, rate limits, new frontier models — including capacity calls like disabling a model that consumed disproportionate limits and moving users to a predictable one.

• A governed MCP gateway, verified end to end. I built the corporate MCP gateway that puts internal MCP endpoints behind one authenticated entrance: corporate sign-in through a per-endpoint OAuth 2.1 authorization server, policy authorization evaluated before each individual tool call, per-endpoint tool filters, and audited invocations that, wherever the backend acts on the caller's behalf, execute as the named user with that user's own permissions. The full chain — sign-in, token, policy, audit, tool call against real data — is verified end to end, authorization state survives restarts, and every credential is reproducible from source control. It is built on ToolHive as declarative Kubernetes custom resources under GitOps, and connecting real AI clients surfaced defects in the gateway itself that went upstream rather than into local patches: a series of issues with root causes pinned to the responsible line, one critical defect fixed by the maintainers within an hour of my investigation, and my first merged pull request credited in the ToolHive v0.46.0 release notes.

• Analytics inside the assistant people already use. Instead of standing up another BI tool, corporate analytics went where the questions get asked: an MCP connector over the analytics database, wired into corporate ChatGPT, with an agent skill that encodes the metric definitions. I iterated that skill against the incumbent BI system until the figures matched, shipped it to production, and extended it on real usage — correctness against the system of record first, convenience second.

• A shared runtime for internal agents. Designing and building the shared runtime internal agents will run on rather than one agent at a time: a self-hosted platform where a new assistant is a stored configuration — model, prompt, tools, subagents, per-tool approval gates — instead of a new codebase. It runs an isolated code sandbox (shell access, real repositories, controlled network), cross-conversation semantic memory, and dynamic code-driven subagent dispatch, behind a full quality gate of automated tests and static analysis that every change must pass — and the defects found along the way were filed upstream across the agent stack rather than worked around locally.

• Company-wide AI adoption, run as a product. Led the rollout of corporate coding assistants (Claude Code) across engineering, QA, DevOps, and analyst roles — one-command onboarding, a centralized support and announcements channel, steady-state adoption across 120+ active corporate users — and standardized AI in the SDLC, continuously extending the corporate configuration with capabilities proven with developers on real repositories first, and making installs modular so each user takes only the components they need. Established corporate ChatGPT Business as a grounded assistant by connecting internal knowledge sources through the retrieval backbone and MCP servers I shipped under this role.

• Saying no to AI when AI is the wrong answer. A standing part of the role: more than one initiative that arrived as an AI request shipped as conventional automation, a process fix, or a playbook instead, and purchases have been declined on measured economics. Owning an AI portfolio means knowing where AI does not belong.

• Hiring and technical evaluation for the AI team. Assessed 100+ candidates and applications for AI-engineering roles (CV review → practical tasks → interviews), designing role-specific tasks and review criteria that tie hiring decisions to real production-delivery ability — holding a senior bar for months rather than hiring down, and building an AI-assisted assignment-review aid, used alongside expert judgment rather than instead of it, when candidates started gaming the tasks.

Expanded Role and Contributions

• AI spend, moving under explicit governance. Driving the AI estate's economics toward full governance: per-project cost attribution so every initiative's spend is visible, consolidation of overlapping subscriptions, and an intake path so new tools enter through evaluation rather than accretion. In motion, not finished.

• Corporate knowledge, searchable by default. Proposed — and carried to a company-wide discussion — that corporate information should be searchable, company-owned, and safely accessible by default, for authorized people and agents alike, within defined security constraints. Grounded assistants are only as good as what they are allowed to see.

• Public artifacts from the role. Created and shipped the company's public repository of agent skills for its trading platform (spotware/ctrader-skills).

• Open-source leverage into production. Converted personal R&D into internal leverage: open-source tooling I author feeds directly into the internal agentic workflows, keeping the company aligned with the fast-evolving MCP / agent ecosystem without depending on any single vendor.

For the AI products and platforms delivered under this role — AILA (AI Localisation Assistant), AIR API (shared retrieval backbone), the YouTrack MCP Server, the Claude Code Gateway, the MCP Gateway, the Managed Agent Platform, and the PeopleForce MCP Server — see the Related Projects panel.

This tenure at Spotware Systems is marked by a platform-first posture: delivering AI products that unlock concrete business outcomes while simultaneously building the shared primitives, access model, and organizational practices that let every subsequent AI initiative arrive faster, cheaper, and better grounded.