> **About** > Staff engineer working at the boundary of AI evaluation, agent systems, and applied empirical research. Background in production ML, linguistics (Paris VII), and multi-agent systems. > > Source: https://jasonstiltner.com/about/ --- # Background Verification-centered empirical research, at the boundary of production AI systems and the evaluation that judges them: AI accelerates the experimentation, controlled evaluation decides what survives, and the evaluator itself is tested before it is trusted to decide anything. Research: AI evaluation, multi-agent coordination, continual learning Production: ML systems, cloud infrastructure Staff Engineer at [GSV AI Labs](https://www.charlieiq.ai), the AI division of the private equity firm Greater Sum Ventures, where we build CharlieIQ — an AI operational suite for vertical SaaS companies that hands engineering, marketing, sales, support, and planning work to agents, with a human approving every consequential action. I was the lab's third engineering hire and its first Staff Engineer. That work spans Minions, CharlieIQ's multi-agent engineering team; [TenantText](https://www.tenanttext.com), its AI resident-communication platform for property managers; and Virtual Front Desk. Separately: AI initiatives across Greater Sum Ventures' portfolio companies, including cross-portfolio fraud detection. Before that, Co-Technical Lead in HCA Healthcare's Accelerated Technologies group, shipping production clinical AI (document intelligence, clinical RAG) at the largest US hospital system; and an Advanced RPA Developer in Accenture's Automation Center of Excellence. ## Scope & Impact ### Cross-org influence AI initiatives spanning Greater Sum Ventures' portfolio companies, including cross-portfolio fraud detection. Separately: built an internal user-feedback platform at HCA Healthcare — routing input from users to development and product teams — pulled in across the advanced-technology division by teams that didn't report to the same org. ### Systems at scale Shipped document-intelligence and clinical RAG pipelines in production at HCA Healthcare, the largest US hospital system. ### Technical direction Staff-level platform and architecture decisions on CharlieIQ's agent systems. [Aegis](https://jasonstiltner.com/projects/aegis/) (event-sourced state, policy gateway, verifiable commitments) is the same kind of infrastructure judgment applied to agent durability. ### Ambiguity Healthcare's regulatory and organizational environment doesn't hand you a spec: HIPAA constraints, clinical stakeholders, and HCA Healthcare's scale as an operating environment all had to be reconciled into one architecture, not designed around in isolation. ## Research ### AI evaluation [Evaluator Assurance](https://jasonstiltner.com/projects/evaluator-assurance/): when can a model-based evaluator be trusted to decide that an agent succeeded? Seven assurance mechanisms preregistered and tested against two corpora of real agentic trajectories. Most failed or weakened on fresh data; the clearest positive result was an evidence-gap rule, where evaluator error ran 25.6% on the cases it fired against 14.4% on the cases it did not (held-out arm, n=1,260). ### Multi-agent coordination [Grounded Commitment Learning](https://jasonstiltner.com/projects/grounded-commitment-learning/): verifiable behavioral contracts for AI coordination. Applies Hart-Moore incomplete contract theory. 40.4% hold-up reduction, [re-derived by CI on every push](https://jasonstiltner.com/corpus/reproducibility/#hart-moore). ### Continual learning [Collaborative Nested Learning](https://jasonstiltner.com/projects/collaborative-nested-learning/): multi-timescale optimization with non-adjacent knowledge bridges. +89% accuracy at high regularization. Pareto-dominant across retention-accuracy tradeoff. ## Methods Claims move through a five-step progression — empirical research, mechanistic validation, formal foundations, executable research, production systems — where each step is what the previous one has to survive before a result is accepted. AI participates freely in the first step and is trusted in none of the rest. The full version, with what currently backs each step, is on the [homepage](https://jasonstiltner.com/#methods). ## Production Experience ### [Document understanding pipeline](https://jasonstiltner.com/projects/document-understanding/) Multi-provider LLM routing with confidence-based escalation. HIPAA-compliant. Human-in-the-loop for low-confidence outputs; corrections captured as reward-shaped labels for offline fine-tuning. ### [Mellifera — voice-first field capture](https://jasonstiltner.com/projects/mellifera/) Hands-free inspection logging for beekeepers working in gloves, with LLM-powered NLU turning speech into structured records. The decision-support models it feeds are specified but not yet trained. ## Technical Stack Languages: Python, TypeScript, SQL ML: PyTorch, TensorFlow, TensorFlow.js Infrastructure: GCP, Terraform, Docker, CI/CD Data: PostgreSQL, BigQuery, vector databases Deployment: Edge, cloud, hybrid ## Education The path here runs through language, not code: six years in Paris working in a second language led to coaching French executives on American business communication and founding a tutoring company, which led to automation — RPA development at Accenture's Automation Center of Excellence (among the first US-based Advanced RPA Developers), alongside software engineering at Atiba Software — and eventually to healthcare-scale ML. ### M.A. Université de Paris VII — Denis Diderot (French-language program) Littérature, Langues, et Civilisations des Pays Anglophones ## Research Interests Pre-representational computation—the operations that exist before and enable representation (whether human language, embeddings, or any other representational layer). - Which compositions of projection, attention, and regularization preserve structure vs. destroy it? - How do systems learn priors over which operations to apply when? - Verifiable behavior grounded in observable actions (scalable oversight) - Alignment for systems whose operations are themselves learned ## Contact [jason@jasonstiltner.com](mailto:jason@jasonstiltner.com) · [GitHub](https://github.com/jstiltner) · [LinkedIn](https://linkedin.com/in/jasonlstiltner) · [Resume available on request](mailto:jason@jasonstiltner.com?subject=Resume%20request)