Matthew Brown

Brahman Command Center

Multi-agent LLM platform · Brahman Capital Corp · sole engineer · 2026

Eight specialized agents that monitor an institutional portfolio, screen for investment ideas, draft morning intelligence, and produce trade decisions on a fixed daily schedule. The system runs unattended on live market data with automatic recovery.

Scale

~22,000 lines of Python across 79 modules

Architecture

8 agents, each an independent HTTP service

Cadence

15-minute intervals through daily scheduled runs

Inference cost

Approximately $2–5 per day at rest

Each agent runs with a dedicated system prompt and a narrow task scope: SENTINEL for portfolio risk, ANALYST for fundamentals, SCOUT for screening, BRIEFING for overnight synthesis, EXECUTOR for trade decisions, plus FLOW and FILINGS for options activity and SEC events, all orchestrated by JARVIS.

The central design decision is the agent collaboration protocol. Rather than hardcoded chains, each agent publishes a capability registry declaring what it can answer and which data sources back it, and any agent can issue a structured query against another. When SENTINEL detects a volume spike, it asks FLOW whether options flow confirms institutional exit, ANALYST whether the move is thesis-breaking, and FILINGS whether insiders have been selling — in parallel — then synthesizes the responses into a single alert. Every collaboration is archived, so any decision can be traced back through the reasoning that produced it.

Market data resolves through a four-tier chain: fresh cache under a fifteen-minute TTL, then the live provider, then stale cache, then a secondary provider as last resort. Each response carries a flag identifying which tier answered, so the system never presents stale data as live. The model layer follows the same principle: rather than pinning a version string that breaks on deprecation, the client probes a candidate list newest-first and caches the first model that responds, distinguishing a genuine not_found from a rate limit so a throttled but valid model is not discarded. A watchdog daemon polls every ten seconds and restarts any failed agent.

Client Intelligence Workspace

Applied agent system · commercial real estate client · 2026

A lead-intelligence platform that scans the market for lease expirations, scores accounts by urgency, drafts outreach, and tracks pipeline. Deployed and running on the client's infrastructure.

The organizing signal is the lease clock. A property whose lease expires in fourteen months is a conversation worth starting now, and the system surfaces it on that schedule rather than waiting for a broker to go looking. Scanning, scoring, outreach drafting, and pipeline tracking sit behind a single workspace in place of four disconnected tools.

Interface detail and usage metrics in preparation.

Architecture vs. Prompting: A Blind Comparison and Fact Audit

Evaluation design · IST 700, Understanding LLMs · 2026

A controlled comparison of a multi-agent system against a single-prompt model on three identical institutional research tasks, followed by a manual audit classifying every specific claim in every response as verified against a live source, hedged, or fabricated.

Claim-level classification across all three tasks. Fabrications included an invented geopolitical event, five fabricated average-volume figures, and days-to-exit calculations derived from them.

On the volume-anomaly task, the multi-agent system identified a position at 12.17 times average volume computed from the live feed, flagged two secondary movers with supporting figures, recommended reviewing recent 8-K filings, and hedged explicitly where a figure was estimated rather than measured. The single-prompt model did not identify which position held the anomaly, invented a geopolitical event to explain volume it cannot observe, produced an authoritative average-volume table in which every figure was fabricated, and carried the invented event into the following task as established context.

The applied conclusion: LLM-as-judge remains appropriate for style, format compliance, and relative ranking on tasks a judge can check. Factual grounding in a consequential domain requires a verifiable-claims harness instead — extract atomic claims, resolve each against a source of truth, and score groundedness rather than persuasiveness.

Pre-Action Phishing and BEC Triage

Applied NLP · IST 664, Natural Language Processing · sole author · 2026

A text classifier for the operational decision in business email compromise — not whether a message is spam, but whether a human should review it before an employee acts on it — built around the asymmetric cost of the two error types. I chose the problem because I have been on the receiving end of one of these attacks, and the failure is never that the message looks like junk. It is that it looks entirely ordinary.

Corpus

3,000 messages, class-balanced

Validation

Five-fold cross-validation

Best result

F1 0.974, logistic regression

Cost model

10:1 false negative to false positive

Tuned threshold

0.25, against a default of 0.50

The pipeline runs from tokenization through cross-validated evaluation across three feature families: unigram presence as a baseline, bigrams, and an original manipulation-marker feature counting the social-engineering cues phishing depends on — urgency, money requests, authority impersonation, credential bait, capitalization intensity, and a length-normalized pressure ratio. Trained in isolation, the manipulation markers reach 0.678 accuracy against a 0.50 chance baseline, confirming independent signal. Logistic regression over the hand-built feature matrix reached F1 0.974, and its calibrated probabilities made cost-based threshold tuning possible.

Does Shorting IPOs Into Lockup Expiry Generate Alpha?

Quantitative research · Brahman Capital Corp · 2026

A benchmark-corrected test of a widely held trade across 61 U.S. operating-company IPOs from 2016 to 2024, spanning four holding windows, including a formal retraction of the prior draft's headline result.

The universe deliberately includes known underperformers to avoid survivorship bias, with SPACs and closed-end funds excluded by construction; of 73 candidates, 61 had clean IPO dates and price coverage, and the remainder were dropped rather than estimated. Tested broadly, the trade does not work — no window and no benchmark produced significant short alpha.

Portfolio AI — Transparency Demo

Interface prototype · IST 675 · team project · 2026

A comparative prototype contrasting how financial AI presents allocation advice today against an interface that surfaces its own confidence, sources, and conditions of validity.

The standard output gives an allocation and general reassurance about proprietary methodology. The redesigned output pairs the same allocation with an explicit confidence score, plausible ranges per holding, dated sources, the calculation method, and a list of the conditions under which the recommendation stops holding. The team's argument is that the polish intended to make AI advice feel trustworthy is the same design choice that makes its quality impossible to assess before acting.

View the live prototype