I lead global alliances with frontier AI labs, hyperscalers, data platforms, and enterprise SaaS companies, building the partnership operating models that turn ecosystems into measurable growth.
The same questions shape the agentic workflows and operating systems I build for myself: how to investigate difficult problems, support judgment, produce finished work, and carry what one run learns into the next.
Most have a user base of one. I describe the machinery here, not the material that runs through it. The work supplies problems at scale; personal systems make them easier to isolate and test. What works moves in both directions, improving the next run and the larger operating model.
-
Research to roadmap — A repeatable, agent-assisted pipeline bringing together scholarly research, competitive intelligence, GitHub implementation patterns, market data, and captured social signals. Observation stays separate from inference; the result becomes capability boundaries, testable questions, stage gates, and build roadmaps.
-
Genealogical and historical research — A source-governed environment with more than 600 canonical person records, tracing one Armenian family from seventeenth-century New Julfa, through the commercial networks its merchants built across Asia, to the present day. Claims enter only through explicit, auditable evidence and confidence rules. Successful findings, rejected candidates, blocked paths, and failed routes are logged alike, so agents do not repeat dead-end research. GEDCOM, registers, and dependency views are generated from the canonical record.
-
Model Council — A reusable deliberation function used across projects when a problem benefits from multiple model perspectives. Its form varies by task: independent research, adversarial review, structured debate, blinded judging, consensus checking, or synthesis. An explicit adjudication layer preserves disagreement and records what survives and why.
-
Artifact production — A semantic router across several thousand indexed visual structures, paired with a gated brief, deck, and template workbench. It searches by communication job, verifies candidate fit, and can decline a match. A broader publishing engine with source and claim ledgers, independent vetoes, and reproducible render gates is in development.
-
Agent operations — A least-privilege layer for supervisory agents, intra-family model selection and routing, and bounded research swarms using open-source or long-context models. Fast workers map, extract, and cluster; frontier models validate consequential claims and adjudicate disagreements. Supervisors observe only the state required to do their job.
-
Classical Armenian text workbench, in design — A research, reading, translation, and edition environment for a language poorly served by mainstream tooling. It takes difficult pages through source and witness control, diplomatic transcription, reversible normalization, linguistic analysis, independent translation paths, deterministic checks for omissions, names, numbers, and negation, human adjudication, correction history, and citable editions.
-
Operating continuity — Private tooling makes selected prior work retrievable and carries my decisions, follow-ups, and operating lessons from one agent session into the next. Routine handling is deterministic; agents enter only when judgment is needed.
-
CouncilLogic — The public-alpha tool expression of Model Council: a local-first Python implementation in which heterogeneous providers propose independently, judge blinded candidates, contribute to deterministic aggregation, and produce a final synthesis. Durable local run records and explicit governance boundaries keep the process inspectable; consequential judgment remains with the user.
-
PageParcel — A local-first browser capture extension that turns a page into portable screenshots, structured Markdown, source metadata, and verifiable hashes. Page processing stays inspectable and local by design.