Adam Benenson

Ambient AI layer for the browser that resurfaces relevant information before you search. The problem is structural: knowledge workers lose hours daily not to hard problems but to context reconstruction — finding the thing you already found, rebuilding the thread you already held. Most AI tools respond to prompts; Forethink anticipates context, proactively surfacing connections across your entire digital history. All processing happens on-device, with no data extraction or centralized indexing. Privacy isn't a policy — it's the substrate.

A permissioned research-intelligence platform for professional services firms — domain ontologies spanning legal matters and investment research (firms, strategies, factors, themes). Captures and reasons over a firm's research, matter, and client history through Email/Drive change-data-capture into a temporal-hypergraph knowledge graph with tri-temporal indexing (event, effective, ingest time). Specialized agentic workflows for retrieval, drafting, compliance, and escalation sit on top. Every retrieval and agent action passes through scoped ACLs with information-barrier enforcement and a full audit plane.

A personal digital twin — built to reconstruct and extend my own causal reasoning (synthesizing Pearl's causal hierarchy with Rubin's potential outcomes), belief structure, and intellectual trajectory. Ingests arbitrary data: conversations, documents, reading history, behavioral traces, structured profiles. Extracts and consolidates claims across ten psychological dimensions; maintains a 24,000-episode store, a 50-node calibrated belief graph, and a queryable knowledge graph. Context engineering is formalized as a Markov Decision Process (DT-MDP) with a contrastive IRL retrieval policy. The system runs continuously, improving its own models between sessions. It also orchestrates structured agentic execution across the lab's other systems — augmenting my day-to-day work and accelerating Forethink's development end-to-end.

Compile governance, don't narrate it. Agentic AI rules should produce typed semantic kernels whose obligations are checked, not evaluated like behavior. The forcing move: reached scarce peer-relative decisive stages cannot jointly satisfy consistency, solidarity, and cross-claimant monotonicity (Arrow-Young-Thomson tradition). The kernel target is a product object — five runtime obligations (certifiable, governance-observable, corrigible, compositionally safe, non-vacuous), social-choice graph diagnostics, a spectral scaling witness, an explicit bridge. A machine-checked Lean 4 development paired with a Rust audit pipeline; end-to-end source walks of five deployed agent harnesses (incl. Codex, Claude Agent SDK, OpenClaw) produce machine-checked rejection certificates that surface real classifier failures rather than theoretical possibilities. Zero sorry, admit, or first-party axiom under the verification gate. A chatbot can be behaviorally aligned. An institution has to be legitimate.

Three layered uniqueness theorems for capacity-achieving input priors on Markov kernels: finite-alphabet, compact-Polish discharged, and measure-theoretic packaging.

The finite Čencov–Petz uniqueness theorem: any continuous monotone metric family on the probability simplex is a scalar multiple of the Fisher information metric. Proved via splitting invariance at rational points + density argument.

The spectral theorem for compact self-adjoint operators on Hilbert spaces: a Hilbert basis of eigenvectors exists.

Rellich–Kondrachov compact embedding: H¹ → L² is compact on compact Riemannian manifolds. Euclidean heart via Fréchet–Kolmogorov; manifold glue via finite chart decomposition.

Led the product, engineering, data, and design departments at an enterprise analytics company serving hedge funds, sovereign wealth funds, pensions, and family offices — 140+ institutional investors managing $3T+ in assets. Full org ownership from strategy through execution; member of the executive team.

Analyst on a $2B+ systematic equity book — Sharpe above 2 sustained through the 2008–2009 regime shift. Promoted to PM and managed own $200M+ portfolio in parallel.