This portfolio contains AI systems built for the investment process, covering knowledge management, research and execution, position and risk monitoring, and the productivity and workforce tools that support them. All were designed and built while completing an Executive MBA and working full time as a management consultant, and all are working software in active use.
Overview
Division of Labor to Maximize AI Value for Human Work
Value emerges from developing AI-powered environments in which each task's tool, configured to its own purpose, is coordinated to target the AI at what it does best and the person at their highest and best use.
The machine's share
Unlock Constraints
The repetitive, constraining tasks that crowd out meaningful work are the same tasks AI handles well. Properly integrated, AI can handle tedious and verifiable tasks, organize and exploit large volumes of information, enhance human context, and apply the human's judgment faithfully across time and workflows.
The human's share
Apply Judgment
By providing insight, targeting the AI based on expertise, and directing the AI to expand the working context, humans are free to focus on high-value decisions, provide judgment, and control workflows. The human gains greater capacity to improve the quality, joy, and value of work.
The dividend
Reinvest Time & Focus
Redesigning workflows to target AI at constraints and limitations, while focusing human attention on tasks that require judgment, yields value to workers and the enterprise. This dividend compounds when reinvested in deeper questions rather than solely greater task volume.
01
Knowledge Management & Exploitation
Proprietary data and analyses deliver edge when they can be quickly and reliably connected, deployed, and exploited.
1.1
Agentic Knowledge Management
Every insight, claim, and datapoint connected for internal consistency
Information debt accumulates when insights detach from their evidence and remain unconnected to related insights across an enterprise's body of work. Proprietary data and information's ability to create edge depends on reliable, auditable, and rapid connection to underpin load-bearing decisions.
Solution approach
Design an agent that puts analysts in command of an entire corpus of previously disjointed data and information, ingesting sources without loss, anchoring every insight to exact spans, and auditing the growing body of information continuously.
AI-Native Application
CARLA · Comprehensive Academic Research Librarian Agent
CARLA allows the analyst to exploit a diverse data and information corpus through a single interface. Originally designed as an academic research librarian, CARLA runs on an ontology of works, claims, entities, datapoints, and concepts within a private corpus. Using that map, every claim is bound to the exact span of source text that supports it, confidence is capped by the fidelity of the document beneath it, and both rules are enforced by validators.
5Format-specific, cell-level anchor schemes
12Claim-drift detection modes
6Source tiers
12Provenance states
51Metadata fields
Figure 1 · CARLA's cross-project knowledge map displaying a single-user corpus with the corroboration and tension links between works and an authored concept opened in detail view. Source: system capture as of 2026-07-14.
Beyond the deskThe practical result for an investment process is that any answer, memo, or citation traces back to material the analyst holds, and CARLA reports insufficient evidence instead of closing the gap itself. This allows source discipline to move beyond compliance to something that compounds research and proprietary data value.
1.2
Field & Source Intelligence
Acquiring and incorporating source material
Filings, market data, licensed databases, and field conversations all have to arrive clean, compliant, and ready to support quality research.
Solution approach
Package acquisition as repeatable and automated workflows that compliantly pull from primary sources and create due diligence prep materials that arm management access visits with specific, sourced questions.
Automated AI-Workflows
Research Acquisition Toolkit.Automated acquisition workflows for SEC EDGAR, WRDS, and LSEG Workspace that pull source material compliantly (declared user-agents, rate limits, licensing respected) and land it in analysis-ready form for research workflows.
Executive Meeting Prep.Automated briefing prep package research and assembly for meetings with company executives. Disambiguated entities, tiered sources, and questions tailored to each host, available in desktop and mobile formats. First deployed for May 2026 company visits in Singapore.
Beyond the deskThe corpus is only as good as the information that makes it in, and these are a few of the reliable intake valves.
02
Research & Execution
These tools aid analysts in converting information into insights, theses, and positions with explicit, gated, and repeatable processes that rigorously audit quality and evaluate risk.
2.1
Analytical Depth & Speed
From thesis to risk-aware position
Developing strong trading positions requires highly reliable processes to price assets, read consensus, manage risk, and stress test potential positions before committing capital.
Solution approach
Encode analyst tasks into a trade development pipeline that automates low-judgment tasks, incorporates contrarian review and adversarial testing, and gates every phase on human decisions.
AI-Native Workflow
Long/Short Research System
A human-gated analytical workflow implemented as composable Claude Code skill packages. Designed to carry a long/short thesis from research through screening, valuation, instrumentation, scenarios, sizing, memo assembly, adversarial review, and live monitoring. Implemented and refined in producing an IC-grade memo on a $29M long/short equities and credit basket targeting mispricing around AI capital expenditures.
14Research phases
19Claude skills
21Human decision gates
2Agentic modeling script auditors
Figure 2 · The 24-scenario basket P&L surface from the workflow's convexity phase. Source: project model outputs, single names generalized.
Beyond the deskThis system is designed to improve and accelerate human analyst judgment by removing or minimizing current constraints and limitations on analysts' ability to incorporate data and adversarial contentions into analysis. Its automated process logging features help measure analytical process quality and ensure traceability and auditability.
2.2
Analytical Edge
Operational insights to drive the trade
Feeding the above analytical workflow and its frameworks with the right data and insights requires asking the operational questions that challenge consensus market assumptions.
Solution approach
Develop information pipelines and analytical frameworks to understand sectors, industries, firms, and leading signals to develop nuanced foresight.
AI-Native Workflows
Sector Strategic Screen.A screening matrix that maps sector participants to value chain positions and strategic posture, built to select coverage candidates before deep work begins.
Industry Supply-Demand Analysis Workflow.A workflow to normalize and evaluate supply and demand signals across currency, fiscal year, and other confounding factors to differentiate between durable demand and artificial demand created by hoarding, forward buying, and other market inefficiencies.
Firm Strategic Analysis Workflow.Distributable plug-in that ingests filings from SEC EDGAR and institutional data sources to evaluate a given firm's strategic position within its industry's strategic group. Strategic alternative scoring and forward-looking analysis is anchored in a quantitative claim source ledger carrying as-of-date and confidence markers for auditability in conducting due diligence and preparing for management engagement.
Deal Signal Pipeline.Alternative data development workflow to identify and evaluate acquirer-target pairings, which draws on public and licensed-for-use information with built-in compliance controls.
Figure 3 · SPACE matrix directional vector developed during an analysis of Cheniere Energy using the firm strategic analysis workflow in June 2026.
Beyond the deskThese tools are highly flexible plug-ins and skill packages that react to analyst judgment and direction to add depth and speed to analyses on consensus-challenging questions.
03
Position & Risk Monitoring
Sizing, scenario forecast, and monitoring workflows allow the analyst to continuously understand the position's risk given changes in the market.
3.1
Managing the Trade
Keeping the position tied to its rationale
A locked memo is a set of claims about the future, and the claims age. Sizing, instrumentation, and timing criteria in the memo must be translated into trade monitoring mechanisms.
Solution approach
Build scenario revaluation and position monitoring processes to surface timely and disciplined risk control, stop loss, profit taking, or other disposition decisions.
AI-Native Position & Risk Toolkit
Trade Instrumentation.Skill packages within the Long/Short Research System (Research & Execution) to select the most appropriate trade expression and structure per name and pair based on the thesis and research.
Position Sizing & Convexity Skills.A gated workflow to develop positions within a portfolio, leveraging composable convexity and scenario skill packages to revalue the book under structured scenarios and surface sensitivities to understand risk before finalizing sizing.
Memo Monitoring.Automated P&L monitoring with alerts for catalyst indicators, custom thresholds, and both profit taking and kill criteria to force intentional dispositions over the planned hold period.
Deal Signals Monitoring Layer.Skill packages within the Deal Signal Pipeline (Research & Execution) to track signal changes per name and per pair over time.
Figure 4 · P&L monitoring from Long/Short Research System's first application to produce an IC-grade memo on a $29M long/short equities and credit basket targeting mispricing around AI capital expenditures.
Beyond the deskMonitoring that forces explicit dispositions keeps a position tied to its rationale after capital is put at risk.
04
Productivity & Administration
These tools convert recurring administrative processes into automated, guided workflows that return time and attention while leaving every decision and submission with the person.
4.1
Guided Administration
Less time navigating business processes
Moving quickly through procurement, sizing, and budgeting processes requires a fluency that the people who visit them occasionally rarely have time to build.
Solution approach
Build guided, single-purpose tools that return a personalized path through each process, making assumptions and confidence visible while leaving submission and approval decisions with the person.
AI-Powered Application
PAM · Procurement Assistance Manager
A multi-turn AI advisor that creates a custom procurement checklist and
drafts required forms, navigating a complex decision tree with disparate
governing regulations in a fraction of the time previously required for
infrequent users. PAM uses a three-tier confidence system that keeps the
model from presenting ambiguity as certainty, and was deployed with
graduate business program administrators.
Developed & Maintained with AI Assistance
MAPS Project Sizing Tool
A model living in a self-contained .html file with intuitive parameter
interfaces, the tool estimates level of effort required for analytics
engagements. Multiple condition axes adjust role-based effort across a
delivery team, providing level-of-effort ranges for specific staff roles.
When a time or staffing constraint cannot be met, the model says so and
ranks which conditions would unlock feasibility. It ships with a written
methods-and-assumptions document to allow teams to recreate the tool to
suit their specific ways of working if needed.
Beyond the deskThese tools quickly guide staff through time-consuming tasks, isolating decisions while automating navigation, search, transcription, and other low-value components of the tasks to recover time.
4.2
Agentic Operations
The tasks that quietly eat the week
Individually each of these chores is relatively minor, but in aggregate, they pull time and attention away from higher-value productive tasks.
Solution approach
Develop a shareable workflow to automate recurring chores for the workforce.
AI-Native Operations Workflows
Course Operations & Weekly Scan.Automated course-operations workflows that upload files, build folders and lessons, and retrieve content through a university learning-management system, replacing one-file-at-a-time clicking. A scheduled scan runs unattended each week, crawling courses, extracting deadlines, detecting changes, and producing a styled report, including a failure report if a run degrades.
Team-Room Booking.Automated reservation workflows that reserve, list, and cancel team rooms by driving the campus booking system directly, chunking longer sessions in compliance with the booking limit and remembering the details that otherwise get re-entered every time. A companion maintenance skill diagnoses and repairs the integration when the upstream system changes.
Duplicate Cleanup.A three-step, fully reversible workflow (inventory, review spreadsheet, recycle-bin deletion) that finds and removes duplicate files based on metadata characteristics, so reclaiming disk space never risks losing the only copy and no longer requires focused user time.
AI-Powered Applications
ReadQueue.An AI-powered reading-queue application that captures articles and research and serves them back as a prioritized queue, with model calls kept behind a serverless proxy.
Model-Discovery & Fallback Library.A compact library that keeps AI-powered applications current as models ship, it discovers the organization's available models, ranks them, calls the best, and steps down the tier on rate limits. It powers both the procurement assistant (above) and the simulation engine (below).
Beyond the deskWeeks of unattended runs are the operational evidence that agentic workflows can hold routine ground and return time and attention back to high-value human tasks.
4.3
Document Production
Documents that assemble themselves
Formatting, citation plumbing, and consistency checks consume hours better allocated to ideation and judgment, and conducting these tasks manually invites errors that can undermine confidence in otherwise excellent work.
Solution approach
Assemble documents from content sources through gated build scripts, preserving canon and verifying parity, so only judgment edits remain by the time a person touches the file.
AI-Native Document Pipelines
Memo & Report Assembly Pipelines.Two pipelines that assemble branded, citation-correct documents from the user's own judgment work in phased and gated workflows within the AI application. The pipelines preserve a user's previously approved document formatting and branding and conduct parity checks with every revision and rebuild against the user's canon before it ships.
Beyond the deskThis site was built using this pipeline, pulling from the volumes of detailed decision work done to create the tools the site showcases.
05
Workforce Development
These tools are a model for upskilling staff out of repetitive, rule-based, and algorithmic roles and preparing them to exercise sound judgment in managing AI capabilities.
5.1
Expertise Engines
Practice that builds and measures mastery
Market judgment and heuristics rest on fundamentals that erode without relevant and repeated practical application.
Solution approach
Engineer engaging daily practice to build durable expertise and measure reliable indicators of mastery.
Developed & Maintained with AI Assistance
LEDGER · Economics Expertise Engine
An installable practice application that builds economics expertise in daily scored sessions, from the national accounts to the Fed's reaction function. Its forecast desk takes standing probability forecasts on real upcoming releases and Brier-scores them when the data prints.
Calibrated Learning Engine.A learning and assessment system for professional certification study. Every item is scored on a correct-by-confident grid and on reveal-to-answer response time, so leaning on test-taking skill shows up as something other than mastery, and the next session drills what the grid says is weakest.
GLYPH · Formal Logic Trainer.The same engine applied to formal logic, from informal argument analysis through graduate-level material. Every puzzle is machine-verified as solvable before it ships, and boards build progressively through six chapters. Live as an installable app.
Figure 5 · LEDGER's forecast desk: the rolling calibration curve across 24 resolved forecasts, and a CPI forecast scored on resolution against the published print. Source: application captures as of 2026-08-26.
Beyond the deskThe engines are built and maintained with AI assistance but scored with deterministic checks in the learning engine's "trust architecture."
5.2
Learning by Simulation
A case that argues back
Cases and slides can describe a decision, but they rarely place a student inside one. Running the review, spending a limited budget on evidence, and defending a recommendation under pushback builds judgment that reading cannot.
Solution approach
Encode facts and extensive context into API calls to the enterprise AI models to create case-based simulations that react dynamically to learner inputs and decisions.
AI-Powered Application
Managerial Economics Simulation Platform
In the working simulation, an antitrust merger review, the student acts as a
Department of Justice analyst: buying evidence against a fixed budget,
computing market concentration, debating the merging parties’ economists
across several rounds, and writing a formal recommendation before receiving
a personalized debrief against a defensible answer. Professors configure
difficulty and structure per run, and prompts and scoring logic stay
server-side. Further simulations, from cartel game theory to surge pricing,
are designed to run on the same engine.
Beyond the deskAny program that teaches judgment under uncertainty, from graduate management to a firm's internal training, can stage its own decisions this way.
The through-line
A tested approach
While each of these systems works and returns time to high value tasks, each can be improved with better access to domain knowledge, market expertise, and proprietary data. This portfolio demonstrates an approach for pursuing ever better insight and judgment in investing.
