writing
-
/01
2026.07.23
METHOD
Inside my agentic workflow: spec, fan-out, verify
How I actually ship with agents. Two shapes, a sequential contract pipeline and a parallel fan-out, both gated by evals. The spec comes before the prompt, the merge happens in plain code, and nothing ships unless the gate is green.
-
/02
2026.07.23
METHOD
Inside the fit judge: an LLM-as-judge that corrected my own labels
An LLM-as-judge scores how well a resume fits a role, gated by a regression suite over a human-labeled golden set. It disagreed with my labels twice and was right both times, caught overqualification that keyword matching cannot, and was non-deterministic at temperature zero until self-consistency voting fixed it.
-
/03
2026.07.02
THESIS
Inside Adoption Chain: an auditable second-order research engine for small-cap equity
Turns a PM's scenario into a consequence map, finds the under-covered small-cap companies on each consequence, and drafts a one-page thesis per name where every claim links to the exact SEC filing sentence. Leverage underneath judgment, never a recommendation.
-
/04
2026.06.24
THESIS
Inside StrikeOut Bot: quantile regression and beat-the-baseline discipline for MLB strikeout prediction
Three LightGBM quantile regressors produce a median and a 90 percent interval, a Ridge stacker blends them against a dumb baseline, and a stratified blind test on unseen post-All-Star-Break data is the only gate that lets a model ship.
-
/05
2026.06.21
THESIS
Inside Study Quest: five architectural decisions in a mastery-gated learning app
A gamified, mastery-gated catch-up learning PWA built for my son. Five decisions: ship for the platform you can deploy from, pick the model on cost, put the integrity boundary in the database, grade only what the model is reliably good at, and make the reward economy a config.
-
/06
2026.06.08
SHIP NOTE
Ship Note: I built a way to watch the watch party
The Knicks are up 2-0 in the Finals and the Game 3 watch party got scattered across three parks. So in one evening I built a way to watch the watch party, using New York's public traffic cameras as a stand-in live feed.
-
/07
2026.06.08
FIELD NOTE
Field Note: turning NYC's public traffic cameras into a live feed
New York DOT puts roughly 957 traffic cameras online as still images with no key and no login. Refresh one on a timer and a still becomes a feed. How I found the right cameras for each location, faked "live" without the page jumping, and handled the cameras that quietly die.
-
/08
2026.05.20
FIELD NOTE
Field Note: Perplexity deep research as the synthesis layer for binary and trinary decision trees
Map the decision as a tree of binary or trinary branches, write a Perplexity deep-research prompt per node, let the model do the cross-source synthesis, then come back to the node and decide. The tree-and-prompt pattern in detail.
-
/09
2026.05.20
FIELD NOTE
Field Note: Discord as the second screen for Claude Code activity
Discord webhooks routed to a private server for ambient AI alerts: ship notifications, build failures, long-task completions, daily summaries. The hook architecture, the channel layout, and what would change at team scale.
-
/10
2026.05.20
FIELD NOTE
Field Note: the Obsidian-Claude Code loop for project tracking and roadmaps
Obsidian as the human-facing planning layer, Claude Code as the execution layer, both reading the same file system. The conventions, where it breaks, and what an AI architect would change at team scale.
-
/11
2026.05.19
FIELD NOTE
Field Note: packaging an emerging sponsor for institutional capital
An emerging sponsor faces a structural problem in the capital stack: senior debt is reachable, LP equity is the wall. The four packaging patterns institutional LPs evaluate against, and the AI-architected workflow that produces the documentation.
-
/12
2026.05.19
FIELD NOTE
Field Note: reading a brokerage flyer through an AI underwriting lens
A brokerage marketing flyer is the highest-volume unstructured document in CRE. Four lines actually matter; the rest is decorative. This Field Note walks the AI-architected workflow that closes the gap between flyer claims and underwritten reality.
-
/13
2026.05.19
FIELD NOTE
Field Note: the NOI projection gap and why hospitality sponsors systematically over-promise
A coastal hospitality asset has an appraisal-stabilized NOI and a sponsor pro forma that differ by 17 percent. Both numbers are produced in good faith. The gap decomposes into four components. This Field Note walks the decomposition and the AI architecture that surfaces it for a credit committee.
-
/14
2026.05.19
FIELD NOTE
Field Note: matching a $25M hospitality deal to five lenders in thirty minutes
Hospitality debt sourcing as constraint satisfaction followed by ranking. The architecture that compresses a week of broker research into thirty minutes of operator review, and what it leaves to the human relationship.
-
/15
2026.05.19
FIELD NOTE
Field Note: stacking R/ECAP, QCT, OZ, and DDD for LIHTC competitive advantage
Five public-records distress designations on a single parcel: QCT, OZ, DDD, R/ECAP, USDA food desert. The Mosaic framework to identify and stack them, and the AI architecture that compresses three weeks of research to three hours.
-
/16
2026.05.19
THESIS
Inside DeckTrainer: a Simply-Piano-style coaching tool for the Hercules DJControl
DeckTrainer is a learn-to-DJ coaching tool for the Hercules DJControl Starlight and Mix. Scope discipline, hardware-constraint surfacing, phased build plan, and a drill engine designed around deterministic scoring functions. The scoping-discipline proof point in the portfolio.
-
/17
2026.05.19
THESIS
Inside the Mosaic feasibility workflow: how I underwrite CRE deals with Claude
The Mosaic feasibility workflow is a Claude prompt suite that produces bank-quality CRE feasibility studies in 90 minutes instead of 30 hours. Standard Underwriting Schema v2.0, dual-scenario NOI, four creative structures, lender probability matching. The eat-your-own-dog-food proof point in the portfolio.
-
/18
2026.05.19
THESIS
Inside TrialEdge: a calibrated probability ensemble for FDA catalyst prediction
TrialEdge predicts FDA PDUFA catalyst outcomes. 4,032 tracked entities, 208 labeled outcomes, AUC 0.686 on the held-out set. The architecture is intentionally small, intentionally calibrated, and intentionally cheap. This post walks the architecture decisions, the eval rig, the cost economics, and what would change starting over.
-
/19
2026.05.19
THESIS
Inside the corridor pipeline: Python automation for LIHTC research
A 15-script Python pipeline that fuses Census ACS, HUD QCT, IRS OZ, USDA Food Atlas, state DDD, and land bank inventory data on a parcel key, scores against the state QAP, and produces an LP-memo-ready output. Five architectural decisions, transferable lessons, free public-data foundation.
-
/20
2026.05.19
THESIS
Inside MES Open: five architectural decisions in an algorithmic trading system
Five architectural decisions in an algorithmic trading system targeting the New York open on Micro E-mini S&P 500 futures. The alternatives rejected, and what each teaches about FDAA judgment in a high-noise low-sample domain where a human discretionary baseline outperforms the model.
-
/21
2026.05.19
THESIS
Inside Marketing Bot v2: five architectural decisions in a multi-pass orchestration system
Five architectural decisions in a multi-pass orchestration system that produces senior-strategist-tier deliverables at software-tier cost. The alternatives rejected, and what each teaches about FDAA judgment in high-quality-bar long-context generation.
-
/22
2026.05.19
THESIS
Inside Atlas: a 19-signal composition scorer for off-market real estate
Atlas surfaces off-market real estate leads by composing 19 distinct signal types across 568,135 parcels in Wake and Durham counties, each signal weighted with an exponential half-life decay. Vision signals from Google Street View and satellite pairs scored by Gemma 3 27B. Total POC spend $54.