I build partnership operating models that turn ecosystems into measurable growth, drawing on experience across technology platforms and global services firms. Today, I lead global Ecosystems & Partnerships at Coforge across frontier AI labs, hyperscalers, data platforms, and enterprise software.
Alongside that work, I build software and research systems for questions I cannot leave alone. The subjects range from AI and markets to history, writing, memory, and everyday product problems. The recurring work is connecting research, judgment, execution, and evidence so the next effort can build on the last.
Most have a user base of one. I describe the machinery here, not the material that runs through it. These are independent projects, with their own scope and stages of development.
-
HKGRO 1854–1941 Jurors List Harness · Archival transcription agent harness · Experimental — Trace connections among people, businesses, and buildings through Hong Kong’s historical jurors lists without quietly rewriting the source. The Python harness compares independent AI readings and sends disagreements through bounded review. Diplomatic transcription preserves spelling, punctuation, and abbreviations as printed. It links decisions to scans, tracks spending, and supports offline replay. Uncertain readings remain visible; search-friendly text stays separate.
-
CouncilLogic Model Council · Heterogeneous model council · Public alpha — Submit a question to AI models from different providers. Models answer independently, then evaluate one another’s responses through blinded judging; CouncilLogic combines their rankings through deterministic aggregation. It preserves an auditable run record, including disagreements; people approve paid runs and record final decisions. Its Python engine runs on your machine, calls cloud model APIs, and provides JSON and MCP interfaces for agents.
-
PageParcel Webpage Capture · Webpage archival & evidence capture · Public alpha — Capture webpages through a Chrome extension, CLI, or MCP server. Bundle screenshots, readable and structured Markdown, source context, and SHA-256 integrity hashes. Annotate captured images to highlight details, explain changes, and exchange visual feedback with people or AI agents. Agents can capture, read, and verify bundles through MCP. Optional on-device OCR extracts text from images. Capture, OCR, annotation, and packaging run locally; PageParcel does not upload captured content for external processing.
-
The Apcar Chronicle · Working research corpus — Reconstruct an Armenian family’s history across centuries of diaspora, from New Julfa to Colonial Oceania, without losing the sources or filling the gaps with plausible stories. Connects people, events, places, and documents through an evidence graph, with research queues, dependency tracking, and publication views.
-
Diplomatic Translation and Transcription Agent Harness · In development — Review historical page scans alongside literal transcriptions and working translations, with uncertain readings flagged against the source image. The first pilot focuses on Grabar (Classical Armenian), testing translation accuracy and consistency in a language with period-specific proper names, complex grammar, and limited machine-translation support compared with modern Armenian.
-
SystemsLog · Persistent memory for people and agents · Working private deployment — Preserve memory, decisions, and unfinished work as conversations end, models change, and agents hand off. A private continuity system that combines durable records, source-linked retrieval, and attributed contributions to recover context and resume work. Human contributions stay distinct from agent interpretations. Corrections enter as timestamped additions; original entries remain exactly as logged.
-
Authoring Studio · Local AI authoring app · In development — Bring book structure, long-form writing, and agent-assisted editing into one Mac workspace. Combines a navigable manuscript map, continuous rich-text editing, structured manuscript import, recoverable revisions, and publication exports. AI editorial changes arrive as reviewable proposals, with the author controlling what enters the manuscript.
-
Impresa Foundry · Domain intelligence and venture discovery · Experimental — Index more than a million domain names, then use contextual retrieval, open-weights models fine-tuned for domain screening, and tiered review to investigate promising names, project fits, and business concepts. Human feedback trains versioned, task-specific open-weights models, while model nominations stay separate from acquisition decisions.
Supporting workflows connect research to build roadmaps, route document and presentation work, coordinate agents, and carry decisions between runs.
Product prototypes and experiments
- Recordbound — A voice-first author-studio pilot connecting audio capture, transcription, source material, and chapter context. Turns recorded conversations into structured book material through consent-aware workflows and editorial review, preserving the connection between what was said, what it supports, and what enters the manuscript.
- Research Desk — A prototype control surface for multi-agent research, connecting question formation, evidence review, bounded investigations, and proposed follow-on inquiries. Designed to make each run inform the next; the interactive prototype demonstrates the workflow ahead of live agent orchestration.
- ReturnLine — A mobile-app prototype that returns a person's own settled intention at a difficult moment, with quiet logging and a history that preserves progress through lapses.
- AI Assessment Lab — A local diagnostic that turns original Azure AI scenario questions, mistakes, and confidence ratings into an inspectable study plan and exportable guide.
- GuidedFishing — A discovery prototype for matching anglers with independent guides, with a voice-agent concierge and assisted inquiry and content operations in the product blueprint. The aim: less administration, more time on the water. Current guide profiles are fictional.
Impresa carries the wider record: professional work, independent systems, earlier ventures, and the questions behind them. LinkedIn carries the career history.