UBS Finance Challenge 2026 — AI Research Intelligence Module

Full-stack AI research tool for the AI Module of the 2026 UBS Finance Challenge.

Pair Trade

• LONG Siemens Energy (ENR.DE) — focus on the Grid Technologies segment
• SHORT Sieyuan Electric (002028.SZ) — Chinese private-sector grid equipment maker

The tool runs a 5-step research pipeline (Tavily search → Anthropic Claude analysis) and produces a structured, exportable research brief that can be embedded directly into the mandatory AI Module slide section.

Stack

Layer	Choice	Why
Search / news	Tavily	1000 free searches/month, structured results with URL + timestamp
LLM	Anthropic Claude (default claude-sonnet-4-6)	Reliable in EEA/Germany; provider is auto-detected from available API keys; swappable to OpenAI / Gemini via LLM_PROVIDER env var
Backend	Python serverless on Vercel	One function (/api/run), 60s max duration
Frontend	Single self-contained index.html	No build step, no React
Hosting	Vercel Hobby Free Tier	$0

Total realistic cost for the entire competition: ~$5–10 (Anthropic API calls; ~$0.20–0.30 per pipeline run with Sonnet, $0.05 with Haiku). Tavily and Vercel are free.

Note on LLM provider: This project initially defaulted to Google Gemini Flash, but Gemini's free tier is not available in EEA / Germany without billing. The current default is Anthropic Claude, which works everywhere and is more reliable for the structured-JSON output the pipeline depends on.

---

Local Setup

git clone <your-repo-url>
cd ubs-ai-module

python3 -m venv .venv
source .venv/bin/activate           # Windows: .venv\Scripts\activate
pip install -r requirements.txt

cp .env.example .env
# Edit .env and paste your TAVILY_API_KEY and ANTHROPIC_API_KEY

Get the keys:

• Tavily (free, 1000 searches/month): https://tavily.com → Settings → API Keys
• Anthropic Claude (free $5 trial credits, ~$0.20–0.30 per run with Sonnet): https://console.anthropic.com/settings/keys

Generate a snapshot

Before the live demo, run the pipeline once locally to produce data/last_snapshot.json. The deployed dashboard uses this as a fallback if the live API call fails (slow venue Wi-Fi, rate limit, etc.).

python fetch_locally.py

This writes data/last_snapshot.json. Commit it; Vercel will serve it as a static fallback.

Local preview

Quickest way to preview index.html locally is just to open it in a browser. The live /api/run won't work locally without a Vercel-style runner, but the snapshot fallback does.

For full local backend testing, use Vercel CLI:

npm i -g vercel
vercel dev

---

Deploy to Vercel

1. Push this repo to GitHub (see below).
2. Go to https://vercel.com → "Add New" → "Project" → import your GitHub repo.
3. Set environment variables in the Vercel project settings:
   • TAVILY_API_KEY
   • GEMINI_API_KEY
   • (optional) LLM_PROVIDER, LLM_MODEL to override defaults.
4. Deploy. Vercel auto-detects:
   • api/run.py → serverless function at /api/run
   • index.html at root → served at /
   • data/last_snapshot.json → served at /data/last_snapshot.json

That's it. Your dashboard will be live at https://<project>.vercel.app.

---

Project Structure

ubs-ai-module/
├── api/
│   ├── run.py                  # Vercel serverless function: POST /api/run
│   └── _lib/
│       ├── data_sources.py     # Tavily adapter
│       ├── llm_client.py       # Provider-agnostic LLM (Anthropic default; auto-detect)
│       ├── prompts.py          # All system prompts + Tavily query catalogue
│       └── research_engine.py  # Pipeline orchestrator
├── data/
│   └── last_snapshot.json      # Cached fallback (committed)
├── index.html                  # Single-file dashboard
├── fetch_locally.py            # Local snapshot generator
├── vercel.json                 # Python runtime config (60s timeout)
├── requirements.txt
├── .env.example
├── .gitignore
└── README.md

---

Pipeline

1. Research Harvest    Tavily search × 17 queries (8 per company + 1 comparison)
2. Sentiment Score     Claude → structured JSON per company
3. Keyword Cluster     Claude → keyword cloud + topic clusters + divergence
4. Thesis Synthesis    Claude → one-line thesis + long/short paragraphs +
                                 key differentiators + why-now + bear cases
5. Compile             ResearchReport with all outputs + sources

Typical pipeline duration: 30–60 seconds end-to-end with Sonnet, 15–25 seconds with Haiku.

---

Configuration

Env var	Purpose	Default
TAVILY_API_KEY	Tavily search API (required)	—
ANTHROPIC_API_KEY	Anthropic Claude API (required unless overriding LLM_PROVIDER)	—
LLM_PROVIDER	anthropic | openai | gemini	auto-detected from available API keys (Anthropic preferred)
LLM_MODEL	Model alias for the chosen provider	claude-sonnet-4-6
OPENAI_API_KEY	Only if LLM_PROVIDER=openai	—
GEMINI_API_KEY	Only if LLM_PROVIDER=gemini (NOTE: not free in EEA / Germany)	—

Vercel deployment env vars

In your Vercel project settings → Environment Variables, add:

Name	Value
TAVILY_API_KEY	your Tavily key
ANTHROPIC_API_KEY	your Anthropic key
LLM_MODEL (recommended)	claude-haiku-4-5 — keeps the pipeline under the 60s function timeout

LLM_PROVIDER is not required — auto-detected from the keys above.

---

Endpoints

Method	Path	Purpose
GET	/	Dashboard (index.html)
POST	/api/run	Run the pipeline; returns full ResearchReport JSON
GET	/data/last_snapshot.json	Static cached fallback

The dashboard always tries POST /api/run first; on failure it falls back to the snapshot and shows a banner.

---

How It Integrates with the Competition

The competition's AI Module section requires teams to use AI to process information beyond manual research capacity, document methodology, and disclose limitations. This tool produces all three:

• Beyond manual capacity — 17 parallel Tavily queries plus four Gemini analyses across both companies in under 60 seconds.
• Methodology disclosure — every step is logged in the dashboard stepper and described in the Markdown export. Source URLs and timestamps are preserved per item.
• Limitation disclosure — the AI Module Slide Notes button copies the exact slide-ready paragraph: what the AI did, what it did not do, where its outputs are weakest.

Slide-ready note (in advance)

The AI module orchestrates a Tavily search-API harvest (~17 queries spanning earnings, segment data, geography, AI/data-center demand, valuation, and policy exposure) and then uses Google Gemini for structured sentiment, keyword clustering, and pair-trade thesis synthesis. It does not access Bloomberg, FactSet, paid terminals, or confidential company information, and does not generate price targets or position-sizing. Sentiment reflects narrative tone in public sources, not market-implied positioning. The AI's role is to compress and structure public research; final pair-trade conviction rests on the team's own fundamental and valuation work.

---

Pre-Presentation Checklist

1. python fetch_locally.py — generate fresh data/last_snapshot.json.
2. Commit and push.
3. Vercel auto-deploys; visit your URL and click Run Analysis to confirm live mode works.
4. Open the production URL in advance from the venue Wi-Fi (or use a phone hotspot as backup).
5. If the venue Wi-Fi is bad, the snapshot fallback kicks in automatically — no panic.