DevPlane
multi-agent kanban + completion-block protocol
Engineering stack
The stack is hub-and-spoke. The discipline is four-S.
Twenty applications under one founder is only feasible with shared substrate: cross-cutting services consumed by multiple verticals, one set of types and tokens, one cadence. The visible apps are the spokes; the hub is what makes the cadence possible.
Hub-and-spoke architecture
Hub: a central registry (people-analytics-toolbox) for service discovery, shared auth, and cross-spoke navigation. Cross-cutting concerns live here once.
Spokes: domain applications that consume hub services and add their own surface — Calculus for metric materialization, Conductor for codegen, Reincarnation for adaptive measurement, AnyComp for compensation decisions, and so on.
Why it matters: every HR product I worked with before kept re-implementing anonymization, metric calculation, segmentation, and survey delivery — and getting each one slightly wrong. Building them once and letting verticals consume them is what makes a single founder productive at this scale.
Custom infrastructure
The pieces I built when no off-the-shelf primitive was good enough. Each is the result of running into the same wall enough times to justify it.
Pill paradigm
typed-transformation flow language
Reincarnation
RID/SID adaptive measurement
Calculus
precomputed metric materialization
Conductor
metadata-grounded SQL/Python codegen
Four-S applied to the practice
Strategy, Science, Statistics, and Systems are the four capabilities I argue need to coexist for analytics to land. They also describe how I work.
Strategy
Every project gets a 'principal issue' framing — the load-bearing decision the work must make legible. Cards on this site lead with that decision, not with stack logos.
Science
Behavioral and decision science as primary inputs. Reincarnation is psychometrics in production. VOI Calculator is decision theory in production. The asymmetry thesis is cognitive science applied to AI.
Statistics
Monte Carlo, regression surrogates, IRT, Bayesian updating — used where they earn their keep. Aggregated dashboards hide variance; segment-grain models recover it.
Systems
Hub-and-spoke architecture, custom flow language (Pills), multi-agent coordination via DevPlane, deterministic-by-default tests. The system makes the science productive.
Standing stack
Runtime
- Next.js 16
- React 19
- TypeScript
- Python 3.13
Data
- Postgres + pgvector
- Supabase
- BigQuery
- Drizzle
AI
- Anthropic API (Claude Opus / Sonnet / Haiku)
- OpenAI API (where lifted)
- Modal (GPU + LoRA training)
- FLUX / Stable Diffusion XL
Infra
- Vercel
- Modal
- GitHub Actions
- Stripe
- Playwright