{"id":216932,"date":"2026-05-04T11:44:17","date_gmt":"2026-05-04T15:44:17","guid":{"rendered":"https:\/\/www.hajim.rochester.edu\/senior-design-day\/?p=216932"},"modified":"2026-05-04T11:44:17","modified_gmt":"2026-05-04T15:44:17","slug":"finch-ai","status":"publish","type":"post","link":"https:\/\/www.hajim.rochester.edu\/senior-design-day\/finch-ai\/","title":{"rendered":"Finch AI"},"content":{"rendered":"\r\n
\r\n\r\n\r\n
<\/div>\r\n\r\n\r\n\r\n
\r\n\r\n\r\n
\r\n\r\n\r\n
\r\n\r\n

Portfolio risk analysis & daily reporting.<\/h2>\r\n\r\n\r\n\r\n

This portfolio risk analysis project was completed as a University of Rochester Data Science Capstone for Senior Design Day<\/a>, with sponsorship from FinchAI.<\/p>\r\n\r\n\r\n\r\n

\r\n\r\n
View a Demo<\/a><\/div>\r\n\r\n<\/div>\r\n\r\n\r\n<\/div>\r\n\r\n\r\n\r\n
\r\n\r\n

We built an AI-powered portfolio risk analysis copilot that ingests daily company news and identifies material risks across a portfolio. It turns scattered news events into a structured daily report that a portfolio manager can review in under 10 minutes. The system uses LangGraph and Claude to classify risks, aggregate evidence, and generate recommendations.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

<\/div>\r\n\r\n\r\n\r\n
\"Portfolio<\/figure>\r\n\r\n\r\n<\/div><\/div>\r\n\r\n\r\n\r\n
<\/div>\r\n\r\n\r\n<\/div><\/div>\r\n\r\n\r\n\r\n\r\n
<\/div>\r\n\r\n\r\n\r\n
\r\n\r\n

The Problem with Manual Portfolio Risk Analysis<\/h2>\r\n\r\n\r\n\r\n
\r\n\r\n\r\n
\r\n\r\n

Portfolio managers tracking dozens of companies face a difficult daily task: reading enough news to identify which holdings face new material risks. Relevant signals are often buried across hundreds of articles, many of which mention a company only in passing.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

Manual review is slow, inconsistent, and impossible to scale. Existing tools either require deep human judgment on every article or miss the nuance needed to separate material risk from noise.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n<\/div>\r\n\r\n<\/div><\/div>\r\n\r\n\r\n\r\n\r\n

<\/div>\r\n\r\n\r\n\r\n
\r\n\r\n

Dataset Snapshot<\/h2>\r\n\r\n\r\n\r\n

FinchAI daily sentiment-event feeds power the portfolio risk analysis system. For production-style reporting, we used the APAC-LATAM portfolio snapshot: 10 daily CSV files from January 31 to February 9, 2026, with 3,620 article-entity rows across 169 unique entities. Each record includes document metadata, company information, business topics, sentiment, salience, and optional stock identifiers. This structure lets the system connect articles to companies, classify risk signals, and generate a portfolio-level report.<\/p>\r\n\r\n\r\n\r\n

\r\n\r\n\r\n
\r\n\r\n

Event Feed<\/h3>\r\n\r\n\r\n\r\n

Each article includes a title, source, URL, entity label, business topics, sentiment class, sentiment score, salience score, and entity summary. These fields let the pipeline classify risk signals without scraping raw article text.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

Development Data<\/h3>\r\n\r\n\r\n\r\n

A larger consolidated dataset of 5,683 observations across 17 columns supported broader development and error analysis, while the portfolio snapshots preserved the daily reporting format expected in production.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

Real-World Gaps<\/h3>\r\n\r\n\r\n\r\n

The feed is realistic and incomplete. About 52% of rows are missing stock identifiers such as ticker, exchange, or MIC code. We kept these rows because the core analysis works at the entity and news-event level, not only at the ticker level.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n<\/div>\r\n\r\n<\/div><\/div>\r\n\r\n\r\n\r\n\r\n

<\/div>\r\n\r\n\r\n\r\n
\r\n\r\n

Pipeline Workflow<\/h2>\r\n\r\n\r\n\r\n

To solve this problem, we built an orchestrator\u2013worker pipeline in LangGraph. The workflow processes each day’s events through five stages.<\/p>\r\n\r\n\r\n\r\n

LLMs handle classification and narrative synthesis. Deterministic Python logic handles severity scoring, category counts, aggregation, and source attribution. This separation makes the system easier to audit and reproduce.<\/p>\r\n\r\n\r\n\r\n

\"Portfolio<\/figure>\r\n\r\n\r\n\r\n
<\/div>\r\n\r\n\r\n\r\n
\r\n\r\n\r\n
\r\n\r\n

01 \u2014 Pull<\/h3>\r\n\r\n\r\n

First, the system loads a named portfolio for a selected date from the FinchAI daily sentiment-event feed. The feed includes company metadata, news events, sentiment scores, salience scores, and business topic tags.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

02 \u2014 Classify<\/h3>\r\n\r\n\r\n

Next, the system fans out across holdings with LangGraph’s Send API. For each holding, it formats event excerpts as TITLE, TOPICS, and CONTENT. It then asks the model to classify risks using a seven-category rubric.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

03 \u2014 Aggregate<\/h3>\r\n\r\n\r\n

Then, deterministic logic merges classification results and groups evidence by company and risk category. The system computes severity by category count: HIGH for two or more flagged categories, MEDIUM for one, and NONE for zero.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n\r\n
\r\n\r\n

04 \u2014 Analyze<\/h3>\r\n\r\n\r\n

After aggregation, an LLM synthesizes the portfolio-level risk picture. It generates an executive summary, per-company key concerns, confidence scores, and recommended actions.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

05 \u2014 Generate<\/h3>\r\n\r\n\r\n

Finally, the system renders a structured HTML report. The report includes company-level risk highlights, severity badges, evidence sentences with source attribution, and a portfolio-wide rollup.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

 <\/p>\r\n\r\n<\/div>\r\n\r\n\r\n<\/div>\r\n\r\n<\/div><\/div>\r\n\r\n\r\n\r\n\r\n

<\/div>\r\n\r\n\r\n\r\n
\r\n\r\n

Risk Taxonomy<\/h2>\r\n\r\n\r\n\r\n

Every flagged event is classified into one or more of seven business-risk categories. The taxonomy is adapted from the Modeling and Detecting Company Risks from News<\/a> framework, which was designed to organize company news into risk categories that are meaningful for financial analysis. We chose this taxonomy because it is broad enough to cover common portfolio risks, but small enough to support consistent annotation, classification, and reporting.<\/p>\r\n\r\n\r\n\r\n

\r\n\r\n\r\n
\r\n\r\n

Supply Chain & Product<\/h3>\r\n\r\n\r\n

Threats to a company’s ability to make, source, or deliver its products and services.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

People & Management<\/h3>\r\n\r\n\r\n

Leadership departures, workforce instability, and internal human capital risk.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

Finance<\/h3>\r\n\r\n\r\n

Financial health, liquidity, access to capital, and earnings risk.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n\r\n
\r\n\r\n

Legal & Regulatory<\/h3>\r\n\r\n\r\n

Lawsuits, regulatory actions, or policy changes that could materially harm the company.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

Macro<\/h3>\r\n\r\n\r\n

Broad external forces beyond company control, including inflation, geopolitics, pandemics, and financial crises.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

Competition<\/h3>\r\n\r\n\r\n

Heightened pressure from rivals or disruptive new entrants to the market.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n\r\n
\r\n\r\n

Markets & Consumers<\/h3>\r\n\r\n\r\n

Revenue threats from weakening demand or shifts in customer behavior.<\/p>\r\n\r\n<\/div>\r\n\r\n\r\n\r\n

\r\n\r\n

 <\/p>\r\n\r\n<\/div>\r\n\r\n\r\n<\/div>\r\n\r\n<\/div><\/div>\r\n\r\n\r\n\r\n\r\n

<\/div>\r\n\r\n\r\n\r\n
\r\n\r\n

What We Built: Project Deliverables<\/h2>\r\n\r\n\r\n\r\n

The final project delivered an end-to-end portfolio risk reporting system with three main components:<\/p>\r\n\r\n\r\n\r\n