Neurons Lab - AI Architect (UA/RU Language speaking)
Requirements
• Four phases — Capture → Connect → Distill → Build — over roughly eight to ten two-week sprints, opening with a fixed-fee two-week Sprint 0 readiness pass (data-access audit, ontology spec, legal checklist across jurisdictions). A family-office workstream runs in parallel on the same squad. • Capture → Connect → Distill → Build • eight to ten two-week sprints • Sprint 0 readiness pass • This is deliberately not a wrapper around an off-the-shelf platform. The client wants infrastructure they own, deployed privately, with role-based access for people and full visibility for the AI. The same architecture becomes a NeuronsLab product line, so you are designing something that has to survive being redeployed for the next client. • NeuronsLab product line • Stage: pre-contract / design-partner negotiation. Duration: multi-phase, ~4–5 months to production for the executive pilot, with rollout beyond it. • Stage • Duration • Reporting: CTO (@Alex Honchar) and CEO are in the room at every key point — architecture, sprint planning, sprint reviews. You own the technical decisions between those points, working alongside an AI Analyst (1.0 FTE) and a Data Engineer (0.5 FTE), plus the client's Head of Security from day one. • Reporting • This role is full-time. • Agentic system architecture end to end: retrieval, tools, orchestration, memory, evals, guardrails • Agentic system architecture • Ontology / knowledge-graph engineering and semantic layers over heterogeneous sources (RDF/OWL, Neo4j, dbt-style modelling — pragmatism over purity) • Ontology / knowledge-graph engineering • RAG / GraphRAG at production quality, including hybrid retrieval and permission-aware retrieval • RAG / GraphRAG • MCP, tool-calling and connector platforms; designing agents that perform actions with side effects safely • MCP, tool-calling and connector platforms • Private / sovereign deployment: VPC, on-prem, self-hosted or open-weight models; AWS and/or GCP data + AI stack • Private / sovereign deployment • Identity, access control and data governance applied to AI systems (RBAC/ABAC, scoping, audit) • Identity, access control and data governance • Strong hands-on Python; comfortable writing the hard 20% of the code yourself • Python • Evals & observability for LLM systems; treating quality as measurable, not anecdotal • Evals & observability • Advanced written and spoken English; can hold an architecture conversation with a CIO and a CISO in the same meeting • English • Knowledge • The current enterprise context-layer landscape — Glean-class platforms, Cohere-class "AI OS" products, Microsoft Copilot / Agents, Gemini Enterprise, Palantir-style foundries — and where each genuinely differs • enterprise context-layer landscape • GDPR and data-residency constraints for multi-jurisdiction European groups; what makes a private deployment defensible • Financial services / private-equity context — investment policy, portfolio reporting, board process — a strong plus • Financial services / private-equity context • OKR / goal-management mechanics, enough to architect for them • 7+ years hands-on AI/ML engineering, of which 2+ years building LLM / agentic systems in production • 7+ years • 2+ years building LLM / agentic systems in production • 3+ years as technical lead or architect on client-facing delivery • 3+ years • Demonstrated ontology / knowledge-graph or semantic-layer work over messy real-world enterprise data • ontology / knowledge-graph or semantic-layer • Experience with regulated or security-sensitive clients (BFSI, government, healthcare) and private deployment • regulated or security-sensitive clients • Experience in consulting or a services business — comfortable being the technical face to a C-level client • consulting or a services business • Comfortable as the most senior technical person on a 2.5-FTE pod, with founders as sparring partners rather than a safety net • most senior technical person on a 2.5-FTE pod
Responsibilities
• Run the Sprint 1 decision spike and write the decision record: one central private-cloud store vs. a semantic layer over the existing systems of record vs. ready platforms (Gemini Enterprise, Glean-class, Cohere-class, open components) — scored on security, access control, speed, cost and reversibility. • Run the Sprint 1 decision spike • Design the ontology / semantic layer for the group: entities, relationships and business definitions spanning people, meetings, decisions, commitments, goals, deals, portfolio companies and documents. • ontology / semantic layer • Architect the connector layer as an execution layer, not just an ingestion layer — MCP / tool-calling (Composio-class or built) so agents can act in HubSpot, mail, Slack and internal systems, not merely read a stream of data. • connector layer as an execution layer, not just an ingestion layer • Design role-scoped retrieval: the principle is that AI sees everything and people keep role-based access. Make that enforceable at the retrieval layer, not just in the UI, and evidence it to the client's security function. • role-scoped retrieval • Architect the agent layer: per-executive skills (Chief of Staff / CIO / CFO / COO), the OKR & drift coach delivered in Slack, and the process miner → optimizer chain. • agent layer • process miner → optimizer • Choose and stand up the private deployment — VPC / on-prem / managed, model selection and routing, cost and latency envelopes. • private deployment • Build the eval and observability harness: correctness, groundedness, access-boundary tests, regression suites before anything reaches an executive. • eval and observability harness • Establish standards and failure-mode design — human-in-the-loop boundaries for agents that take real actions, audit trails, rollback. • standards and failure-mode design • Stay hands-on: implement the critical pieces yourself, review the pod's work, and keep the build portable enough to redeploy as a NeuronsLab offering. • portable enough to redeploy • Explain all of the above to a C-level audience in plain language, in review sessions and working groups. • in plain language
Apply in one click
Upload My Resume
Drop here or click to browse · Tap to choose · PDF, DOCX, DOC, RTF, TXT