Retrieval over EDGAR
- Nightly ingest of 10-K / 10-Q / 8-K filings via EDGAR REST
- Section-aware chunking that preserves Item 1A / MD&A boundaries
- Sentence-transformer embeddings indexed in Postgres pgvector
Group 01
· Capstone · Fall 2025
Stock Investment AI — an educational & paper-trading platform.
A browser-based dashboard that puts learning, real-time charts, rule-based strategy backtesting, and Alpaca paper trading in one interface — with on-demand AI explanations grounded in the exact data the user is looking at, so beginners can move from concept to test to simulated trade without switching tools.
IWhat we built
What problem this solves.
Retail investors are flooded with stock-pick "AI" that recommends with confidence and explains with nothing. The market is full of black-box score generators, paywalled newsletters, and dashboards that lean on past performance as if it were predictive. The information that would actually move an investor's decision — what the company said in its last filing, what it didn't say, what changed since last quarter — sits in dense 10-K and 10-Q documents that almost nobody reads.
Group 01 chose a focused niche: the retail investor who is serious enough to want sources, but doesn't have time to read four hundred pages of SEC prose before placing an order. The product is not a stock picker. It is a reading assistant for filings, grounded in retrieval, that gives a recommendation only when the source documents support one — and refuses, by design, when they don't.
IIHow it works
The system, end to end.
The technical spine is a four-stage pipeline. (1) Ingest pulls the most recent 10-K, 10-Q, and 8-K filings for a watchlist of tickers from SEC EDGAR. (2) Chunk & embed normalizes each filing into semantic sections, embeds with a sentence-transformer, and stores in a vector index. (3) Retrieve answers an investor query by surfacing the three to five passages most relevant to the claim. (4) Reason hands those passages to an LLM with an explicit refusal prompt: if the documents don't support a call, say so.
Two design choices made the difference. First, the system writes its reasoning before its recommendation — a structural commitment that prevents post-hoc rationalization. Second, every claim in the output is anchored to a specific filing paragraph; the UI surfaces those citations inline so the investor can verify the model's reading against the source.
Ingest
Query
ticker · question
Transform
EDGAR retrieval
10-K / 10-Q / 8-K via EDGAR REST
Storage
EDGAR filings
section-aware chunking
Model
Reasoning & refusal
refuse below top-k threshold
Transform
Paper trading loop
Alpaca paper account
Surface
Dashboard
lightweight-charts + explanation
Transform
Evaluation
refusal precision + citation audit
Ingest Transform Model Storage Surface
Feedback
IIIThe stack
What it's built on.
| Retrieval over EDGAR | Nightly ingest of 10-K / 10-Q / 8-K filings via EDGAR REST |
|---|---|
| Reasoning & refusal | Prompt template enforces reasoning-before-recommendation order |
| Paper trading loop | Alpaca paper account for risk-free order routing |
| Evaluation | Adversarial red-team week — graders fed contradictory filings |
Reasoning & refusal
- Prompt template enforces reasoning-before-recommendation order
- Hard refusal path when top-k retrieval scores fall below threshold
- Inline citations rendered as anchors back to source paragraph
Paper trading loop
- Alpaca paper account for risk-free order routing
- lightweight-charts overlay synced with the AI explanation pane
- Strategy backtests run server-side; results cached per ticker / window
Evaluation
- Adversarial red-team week — graders fed contradictory filings
- Refusal precision tracked separately from recommendation accuracy
- Citation-faithfulness audit against a 50-question gold set
IVDeliverables
What the team shipped.
Source repository GitHub · code, tests, README
Demo video Capstone day · screen recording, 4–6 min
Write-up PDF Final brief · methods, evaluation, reflection
Slide deck Capstone presentation · 10 slides
VWhat sets it apart
What sets this capstone apart.
Pick the niche, then build the product.
Not "retail investors." Not "stock pickers." Specifically: the serious retail investor who wants source-cited reasoning over filings, and doesn't have time to read four-hundred-page documents. The whole product follows from naming that user out loud.
Refuse by design, not by accident.
If the retrieved passages don't support a recommendation, the system says so. Refusal is written into the prompt as a first-class response, not bolted on after the fact. That single commitment is what carried the system through the adversarial week.
Every claim, anchored to a filing.
No unattributed scores, no synthetic sentiment aggregation. Every line in the output cites the specific 10-K or 10-Q paragraph it came from — verifiable in one click. The investor reads the model, then reads the source.
VIIInstructor note
How this project landed.
The starting proposal in week eight was broad — "financial literacy and AI-driven stock prediction" — without a specific niche or named competitors. The first review pushed the team to define which investor, against which tool, doing what differently.
They came back two weeks later with the answer that ultimately carried the project: a reading assistant for filings, not a price predictor. That single reframing turned a crowded category into a defendable one and made the technical choices crisp. The capstone shipped on schedule, defended its design against the adversarial red-team without compromise, and stands as one of the strongest pieces of work in the Spring 2026 cohort.
What this project should teach the next cohort: the niche you pick is the project. Pick it carefully, defend it explicitly, and let it dictate every technical choice that follows.