LAURENCE FANG
← RETURN TO HORIZON

04 / HORIZON

Questions still
in motion.

Research directions and engineering notes,
held to their actual evidence.
01

CLOSED RESEARCH

Poker Agent Research

Closed research on game-theoretic decision making in incomplete-information poker.

MECHANISM

Information setsCFRAverage policyBest responseExploitability

OBSERVED OUTPUT

  • A reproducible research environment built around Kuhn Poker and Leduc Poker.
  • Five progressive research notebooks covering fundamentals, CFR, best response, exploitability and Leduc Poker.
  • Scripts for CFR runs, strategy comparison and solver configuration.
A closed research workspace. It does not claim live-game win rates, a commercial poker agent, or a general poker system.
02

LOCAL ENVIRONMENT

Local Open Model Studio

A local environment for LLM inference and image/video model workflows under explicit hardware constraints.

MECHANISM

Qwen3.8-27Bllama.cpp64K contextComfyUISenseNovaMiniMax H3

OBSERVED OUTPUT

  • A local Qwen3.8-27B server with an OpenAI-compatible endpoint and a repeatable health check.
  • Image, image-editing, multi-reference, text-to-video and reference-driven video workflows.
  • Model, cache, output and temporary-file separation, with documented GPU memory scheduling between LLM and generation workloads.
A personal engineering environment, not a hosted inference platform or a benchmark claim about model quality.
03

ENGINEERING NOTE

DeepSeek Harness Engineering

Input-pipeline engineering for general file attachments in an existing AI developer tool.

MECHANISM

fileIdsAttachment stateSerializationsendSessioncreateDraftFiles(files)

OBSERVED OUTPUT

  • Extended the conversation input flow beyond image-only attachments.
  • Carried file identifiers through submission, captured attachment state and cleared consumed files after successful sends.
  • Mapped the change across facade, hub and service layers.
Evidence of work in an unfamiliar codebase; it does not claim an official upstream contribution or merged pull request.
04

IN PROGRESS

Multimodal Financial Research

Exploring text and voice/audio signals in earnings calls and financial communication.

MECHANISM

Earnings callsTextVoice / audioPredictionRisk analysis

OBSERVED OUTPUT

  • Research direction and signal space defined.
  • Potential downstream questions identified around stock-related prediction and risk analysis.
Research in progress. No final experimental results, predictive performance or financial claims are presented.