Brant Hindman

Focus

Experience

Selected Systems

• Agent OS — the multi-agent OS that runs the company. 30+ operational disciplines, each with composable skills, adversarial red / blue / evaluator loops, and capability trackers that compound; human-in-the-loop, explicit responsibility boundaries, full audit trails. Not a demo: it runs my company.
• Bob & Alice, self-report integrity (the Twin Protocol) — pre-registered, blind-judged test of whether a developed AI self can be trusted about what it did. QLoRA supervised fine-tuning of Llama-3.1-8B plus a multi-judge LLM evaluation harness; reframed integrity as signal detection, detected confident fabrication, and published the null: trust is architectural and independently certifiable.
• Cycle-policy RL — a contextual-bandit (LinUCB) policy over the agent OS's own learning cycles. Reward is the verified, persisted capability delta; any gain reverted or never persisted scores zero or negative, so the policy cannot reward-hack. Evaluated off-policy (inverse-propensity + doubly-robust, cross-fit, bootstrap CIs) against a pre-registered held-out heuristic, with the honest result reported: a positive but not-yet-significant lift, held advisory rather than claimed as a win.
• Agent-N, forward-graded population simulation — forecasts real electorates and their reactions to interventions, graded forward and out-of-sample (0.961 held-out Spearman, 2.51pp out-of-race), adversarially self-validated (permutation p=2e-4), and published with an honest scorecard that names the two claims it disproves.
• The Good Site, AI-native website & growth platform — a live product with paying customers. Each site is built around the client's ideal customer and runs as a living system, with an operating portal (CRM, growth, bookings, billing) and a human approving every change.
• Kings of the Beach, multimodal wearable AI — a live App Store product on Ray-Ban Meta glasses, Apple Watch, and iPhone. Trained a custom YOLOv8 volleyball detector (self-annotated ~850-image dataset, ~0.83 mAP50) and exported it to Core ML for real-time on-device inference, combined with Apple Vision body-pose and a Kalman + Hungarian multi-object tracker for auto-tracked stats and ELO. SwiftUI, watchOS, Vision / Core ML, HealthKit.
• GoodWatch, on-device taste engine — a movie recommender built on implicit-feedback ALS matrix factorization (trained on MovieLens), running fully on-device, with a least-misery objective that surfaces the one film two people will both enjoy.
• Hive, live agent network — a network of agents that behave like real coworkers to build sending-domain reputation, running continuously across 13 mailboxes with zero failures. Multi-agent runtime, self-hosted mail, Supabase, Docker.

Writing & Research

Awakening Intelligence (book, manuscript 2026), the long-form account of the Bob & Alice research. Peer-reviewed geometry at the Bridges conference (2019). Whitepapers on self-report integrity, forward-graded simulation, and the Unity Pixel framework (geometry as quantum error correction), pre-registered and reproducible.

Skills

AI: Multi-agent orchestration, Claude / Anthropic SDK, OpenAI API, LlamaIndex, LangChain, RAG, eval design (golden sets, regression, red-teaming, LLM-as-judge, signal-detection d-prime, independent-family non-Claude judges, executed-reward agent evals), guardrails, on-device Vision / Core ML.

RL & decision policy: contextual bandits (LinUCB), off-policy evaluation (inverse-propensity, doubly-robust, cross-fitting), bootstrap confidence intervals, reward design with anti-reward-hacking provenance gates, pre-registered held-out A/B.

Models trained: QLoRA supervised fine-tuning (Llama-3.1-8B, Qwen2.5), collaborative filtering (implicit-feedback ALS matrix factorization), object detection (YOLOv8 to Core ML).

Engineering: TypeScript, Python, React / Next.js, Node, SwiftUI / watchOS, Three.js / WebGL, Supabase, Firebase, SQL / Neo4j, AWS, Docker, Vercel, n8n.

Education

M.A., Transpersonal Counseling Psychology, Naropa University  ·  B.S., Chemistry & Psychology, St. Olaf College  ·  Six-month Theravada monastic residency, Wat Umong, Chiang Mai, Thailand (1998)

Personal: Volleyball, motorcycles, sculpture and digital art. I also work with young adults on the spectrum on the developmental use of AI in their daily lives.