Dual-LLM equity qualitative overlay: Qwen via Alibaba Cloud Model Studio / DashScope vs DeepSeek — side-by-side comparison with agreement scoring, fail-closed LLM handling, evidence provenance, and a human-review gate.
A sanitized, judge-facing vertical slice of the private Pantheon Research production system — cloud deployment, data governance, dual-model comparison, and product-grade UI. Not an API wrapper.
Judges: start with docs/judge_evidence.md for a 3-minute verification path covering Alibaba ECS, DashScope/Qwen, RDS selected mirror, fail-closed model handling, and reproducible local smoke tests.
| 🌐 Live Product | https://pantheon-research.com |
| ☁️ Alibaba Cloud Deployment | http://8.222.191.152 |
| 🎬 Demo Video | https://www.youtube.com/watch?v=68lceOACLKo |
| 🖼️ Deck | Google Slides |
| 💻 Public Code | https://github.com/0xjacobzhao-byte/pantheon-research-qwen-hackathon |
The live product and Alibaba deployment run the full private production system. This repository is the sanitized, self-contained slice judges can clone and run in minutes.
-
Verify Alibaba ECS proof
curl -s http://8.222.191.152/api/proof/alibaba-cloud | jq -
Inspect deployment proof code
backend/app/alibaba_cloud_proof.py -
Inspect actual Qwen / DashScope API call
backend/app/qwen_overlay.py -
Run the offline demo locally
docker compose up --build ./scripts/judge_smoke.sh
-
Read the full evidence guide
docs/judge_evidence.md
This public repository is a sanitized, self-contained hackathon slice. The complete production codebase lives in the private Pantheon Research repository:
https://github.com/0xjacobzhao-byte/Pantheon-Research
It remains closed-source to protect proprietary trading-strategy IP, provider integrations, operational runbooks, and production data infrastructure. Qwen Hackathon judges may request temporary private access from Jacob Zhao for verification.
git clone https://github.com/0xjacobzhao-byte/pantheon-research-qwen-hackathon
cd pantheon-research-qwen-hackathon
docker compose up --build # frontend :5173 · backend :8000
./scripts/judge_smoke.sh # end-to-end smoke test (offline, no secrets)Manual setup (no Docker)
Backend (Python 3.11–3.12):
cd backend && python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
uvicorn main:app --reload --port 8000Frontend (Node.js 18+):
cd frontend && npm install && npm run dev| Service | URL |
|---|---|
| Frontend | http://localhost:5173 |
| Backend API | http://localhost:8000 |
| API Docs (Swagger) | http://localhost:8000/docs |
Verify live Alibaba deployment:
curl -s http://8.222.191.152/api/proof/alibaba-cloud | jq
Pantheon Research is a framework-first, data-governed, human-in-the-loop AI research operating system. This repository showcases the dual-LLM qualitative overlay feature: for a given stock ticker, two independent LLM providers (Qwen and DeepSeek) each produce a structured qualitative assessment, and the system compares them side-by-side.
Strategy ──▶ Information ──▶ Signal ──▶ Trading
| Layer | Role |
|---|---|
| Strategy | Investment thesis and universe selection |
| Information | Evidence pack: quantitative metrics, fundamentals, market data |
| Signal | Dual-LLM qualitative overlay generates structured assessment fields |
| Trading | Human-in-the-loop decision gate (LLMs never execute trades) |
Safety: LLMs do not execute trades. Every signal passes a human-review gate. Pantheon Research is not an autonomous trading bot.
| Capability | Implementation |
|---|---|
Fail-closed model states — missing key → BLOCKED_BY_MISSING_CREDENTIAL, bad JSON → PARSE_ERROR, missing sample → QWEN_NOT_GENERATED |
qwen_overlay.py · models.py |
Evidence hashing — every pack committed to a sha256 content hash threaded into each comparison |
evidence_pack.py |
Dual-model agreement & divergence — two independent models, per-field divergence, data_state (LIVE_DUAL / OFFLINE_SAMPLE / MIXED / PARTIAL / BLOCKED) |
comparison.py |
Human-review gate — low agreement or major divergence flags human_review_required; fail-closed yields NOT_COMPARABLE |
comparison.py · OverlayComparisonPanel.tsx |
Multi-asset scope — Macro · TA · FICC (FI/FX/Commodity) · Equity module grid with per-module data_state |
sample_modules.py · ModuleSnapshotGrid.tsx |
| Research-Ops panel — governance snapshot: provider config, coverage, per-ticker state | data_quality.py · DataQualityPanel.tsx |
| Validation methodology — overlay is a tracked signal, not an alpha oracle | docs/validation_methodology.md |
| Alibaba live deployment proof — host-honest proof endpoint + admin-gated live Qwen smoke | docs/live_proof.md |
| Component | Detail |
|---|---|
| Cloud Provider | Alibaba Cloud |
| AI Provider (Qwen) | Alibaba Cloud DashScope / Model Studio |
| Compute | Dockerized FastAPI behind Nginx on Alibaba Cloud ECS |
| Host Detection | Honest via alibaba_hosted (same image runs on Railway and Alibaba ECS) |
| Database | Alibaba RDS PostgreSQL-compatible — selected evidence mirror |
The /api/proof/alibaba-cloud endpoint returns deployment metadata with booleans only — no keys, tokens, or connection strings. It makes no external calls, so it never claims connectivity it did not verify.
| Proof File | Purpose |
|---|---|
backend/app/alibaba_cloud_proof.py |
Host/runtime, credential state, ECS/RDS/DashScope service map, safe/non-claims |
backend/app/qwen_overlay.py |
Actual Qwen / DashScope API call implementation |
RDS provisioning is kept distinct from full production-data migration:
{
"role": "selected evidence mirror",
"mirror_state": "partial_selected_mirror",
"connected": null,
"production_data_migrated": false,
"full_production_clone_verified": false
}On the live ECS box, RDS is deployed and connected (connected: true). This offline demo performs no probe (connected: null). Full production-data migration is not claimed without core row counts and API read-path verification. See docs/live_proof.md for the captured live response and docs/alibaba_deployment_parity.md for the full breakdown.
| Property | Value |
|---|---|
| Provider | Alibaba Cloud Model Studio / DashScope |
| Base URL | https://dashscope-intl.aliyuncs.com/compatible-mode/v1 |
| Model | qwen-plus (configurable via QWEN_MODEL) |
| Auth | Bearer token (DASHSCOPE_API_KEY) |
| Protocol | OpenAI-compatible chat completions |
Default mode is offline — no API key required. Bundled sample data in data/. Set DEMO_MODE=live + DASHSCOPE_API_KEY for live calls. See docs/qwen_integration.md.
| Metric | Value |
|---|---|
| Qwen comparison-capable | 312 tickers |
| Markets | US 117 / CN 69 / HK 103 / SG 23 |
| Healthy comparisons | 312 / 312 |
| DeepSeek baseline universe | 1,331 |
Full-universe parity was intentionally not pursued for the public demo. The current Qwen coverage prioritizes liquid, judge-relevant equities across US, China, Hong Kong, and Singapore; low-liquidity tail coverage remains private/backlog.
- Select Ticker — Choose MA (Mastercard) or NVDA (NVIDIA)
- Load Evidence — Backend loads quantitative metrics from
data/ - Qwen Overlay — DashScope generates structured assessment
- DeepSeek Overlay — DeepSeek generates independent assessment
- Comparison — Agreement score, tone classification, divergences, evidence gaps
- Human Review Gate — Low agreement or major divergences → human review flagged
Each overlay produces structured fields: business_quality, moat, pricing_power, capital_allocation, red_flags, confidence (0–1), missing_evidence.
Example comparison output (NVDA, offline sample)
{
"ticker": "NVDA",
"data_state": "OFFLINE_SAMPLE",
"agreement_score": 0.44,
"agreement_level": "LOW",
"qwen_tone": "conservative_positive",
"deepseek_tone": "conservative_positive",
"divergences": [{ "field": "pricing_power", "severity": "major" }],
"evidence_gaps": ["No competitive ASIC roadmap analysis"],
"human_review_required": true,
"human_review_reason": "Low agreement between providers."
}data_state is the honest headline: LIVE_DUAL, OFFLINE_SAMPLE, MIXED, PARTIAL, or BLOCKED. When a provider fails closed, the comparison is NOT_COMPARABLE — no agreement score is fabricated.
Full endpoint reference (21 endpoints)
| Method | Path | Description |
|---|---|---|
| GET | / |
Root info |
| GET | /health |
Health check |
| Core | ||
| GET | /api/project |
Project metadata |
| GET | /api/evidence/{ticker} |
Evidence pack + provenance (sha256 content hash) |
| GET | /api/overlay/qwen/{ticker} |
Qwen qualitative overlay |
| GET | /api/overlay/deepseek/{ticker} |
DeepSeek qualitative overlay |
| GET | /api/comparison/{ticker} |
Full dual-provider comparison |
| GET | /api/data-quality |
Research-Ops / governance snapshot |
| GET | /api/modules |
Module snapshot grid |
| GET | /api/validation |
Forward-validation methodology |
| GET | /api/demo-flow |
Demo flow steps |
| Alibaba Proof | ||
| GET | /api/proof/alibaba-cloud |
Deployment proof (v2, canonical) |
| GET | /api/alibaba/proof |
Deployment proof (alias) |
| GET | /api/alibaba/qwen-config |
Qwen / DashScope configuration |
| Production-Feel Panels | ||
| GET | /api/ticker-profile/{ticker} |
Ticker profile with KPI cards |
| GET | /api/ticker-profiles |
List available ticker profiles |
| GET | /api/provider-health |
Provider health snapshot |
| GET | /api/validation-timeline |
Signal lifecycle timeline |
| GET | /api/mini/macro |
Macro regime mini panel (context-only) |
| GET | /api/mini/market-pulse |
Market Pulse / TA mini panel (context-only) |
| GET | /api/mini/ficc |
FICC mini panel (context-only) |
cd backend && python -m pytest # 80 backend tests
cd frontend && npm test -- --run # 9 frontend tests
cd frontend && npm run build # production build
docker compose config # validate compose file
./scripts/judge_smoke.sh # end-to-end smoke| Layer | Technology |
|---|---|
| Backend | FastAPI · Python 3.11–3.12 |
| Frontend | React 18 · TypeScript · Vite 6 |
| LLM (Qwen) | Alibaba Cloud DashScope (OpenAI-compatible) |
| LLM (DeepSeek) | DeepSeek API (OpenAI-compatible) |
| Database | PostgreSQL (Alibaba RDS-compatible) — production only |
| Deploy | Docker Compose · Alibaba ECS (Nginx → FastAPI) |
| Tests | pytest (backend) · vitest + Testing Library (frontend) |
Jacob Zhao — 0xjacobzhao-byte
License: Apache-2.0 — see LICENSE
No API keys, private user data, live trading credentials, production secrets, or private financial records are included in this repository.