02 · Applications

Selected AI and Legal Technology Portfolio

Independently designed and built, on personal time, using public data (2024–Present).

Securities Fraud Analyzer

Maturity
Private pipeline, runs daily

Inputs
Public market & filing data, 5,000+ U.S. issuers

Models
Claude · GPT · Gemini, orchestrated

Tests
900+ files incl. financial-accuracy checks

Securities Fraud Analyzer

I built a daily production pipeline that ingests data on 5,000+ U.S. public issuers, flags potential corrective disclosures from abnormal price movements, and prepares PSLRA event-study damages estimates (abnormal returns, stock turnover, 90-day lookback) and insider-trading analyses, generating lawyer-reviewable assessments of pleading-stage strengths and weaknesses through a triage dashboard.

The system orchestrates multiple frontier models (Claude, GPT, Gemini) and encodes securities-litigation judgment, including loss causation, event-study methodology, and scienter signals, into a repeatable, auditable workflow. A 900+ test-file suite, including financial-accuracy tests for event-study and damages calculations, safeguards litigation-critical outputs.

Habeas Watch

Maturity
Working prototype · pro bono

Inputs
CourtListener / RECAP dockets and filings

Models
Claude + GPT, dual-model extraction

Litigation Data and Docket Intelligence

I built Habeas Watch, a pro bono immigration-habeas platform that ingests CourtListener/RECAP docket data, retrieves court documents, classifies cases with dual-model (Claude + GPT) extraction, and surfaces citation-forward summaries for lawyers handling urgent detention matters.

Privilege-aware legal AI control plane

Maturity
Architecture & Azure prototypes · not deployed at any firm

Integration targets
iManage · NetDocuments · Relativity · SharePoint

Secure, Conflicts- and Privilege-Aware Legal AI Control Plane

I built Azure-first prototypes of a governed law-firm AI control plane for deploying frontier AI tools inside a firm’s own Microsoft tenant. The architecture is designed to integrate advanced AI tools from foundation labs (such as Anthropic’s Claude Cowork and OpenAI’s Codex productivity tools) with existing firm systems, including iManage, NetDocuments, Relativity, and SharePoint.

The project enforces conflicts walls; adds source-grounded retrieval, citation verification, output controls, attorney-review gates, and client/matter budget attribution; and is informed by my published analysis of attorney-client privilege risk in AI workflows. A manifest-driven tools marketplace is intended to empower and leverage lawyer-builders and KM/innovation personnel by allowing them to easily roll out powerful tools that connect to existing data sources within the firm, while enforcing the firm’s risk controls and security boundaries.

Also built

Matter Review and Litigation Knowledge Management. I built an AI-native matter-review workspace prototype with matter vaults, chat, tabular document review, source citations, and Azure-hosted deployment architecture. I developed document-conversion and knowledge-management tooling using Mistral OCR, Tesseract, and Deepgram.
Agentic Engineering. I built the engineering substrate beneath this work: multi-agent orchestration with worktree isolation and permission boundaries, multi-model provider routing (Claude/Codex/Gemini), and a fail-closed Cloudflare/GitHub security control plane for protected deployment approvals and short-lived token brokerage. The design supports unsupervised, long-running agent sessions without permission-prompt interruptions while holding strict security boundaries and a minimal blast radius against supply-chain and other AI-development risks.
AI and Systems Foundations. I built projects to understand modern AI and high-performance systems: fine-tuned open-weight LLMs on a custom corpus (Hugging Face / RunPod GPUs), ran self-hosted model inference locally on GPU, and built a retrieval-augmented generation (RAG) pipeline over a Pinecone vector store to compare fine-tuning against RAG directly; optimized Anthropic’s public performance take-home by roughly 85× over the provided baseline; and wrote a GPU-accelerated Rust solver (linear programming, dynamic programming, and MILP) for an online optimization challenge.