Adversarial multi-agent due-diligence on public-company tickers.
🔗 Live · https://diligence.duckdns.org 📺 Walk-through · link added with hackathon submission 🏆 Track · lablab.ai Milan AI Week '26 — Vertex AI Gemini · Featherless · Speechmatics · Vultr 💸 Cost · ~$1.25 per fresh ticker ($0.21 Gemini + $1.04 Speechmatics enhanced); cached re-runs free
Paste a public-company ticker. Diligence reads the 10-K, 10-Q, and the latest earnings call (audio → diarized transcript), runs five adversarial AI agents that build the bull case and the bear case from the same evidence, then surfaces the disputed facts — ranked by materiality, citation-linked back to the source filing section or transcript timestamp.
Built for the analyst who skips the first eight hours.
Research-grade, not advisory. No buy/sell/hold. Evidence and counter-evidence; the human decides.
Single-LLM summarisers are biased toward the loudest passage. Diligence forces two agents to argue opposite cases from the same primary-source claim IDs, then uses a third agent to score where they materially disagree. Every claim must cite a claim_id from the Filing or Call analyst's output; un-cited assertions are flagged automatically.
┌───────────────────────────┐
10-K + 10-Q ─►│ Filing Analyst (Gemini) ├── claim_ids F-001…
└───────────────────────────┘
┌───────────────────────────┐
Earnings call ─►│ Call Analyst (Gemini) ├── claim_ids C-001…
└───────────────────────────┘
│
┌──────────────┼──────────────┐
▼ ▼
┌────────────────┐ ┌────────────────┐
│ Bull (Qwen3) │ parallel │ Bear (Qwen3) │
└────────────────┘ └────────────────┘
│ │
└──────────────┬──────────────┘
▼
┌───────────────────────────┐
│ Reconciler (Gemini) │── DisputedFacts
│ · diff bull vs bear │ ranked 1–10 by
│ · score materiality │ materiality
│ · flag uncited claims │
└───────────────────────────┘
Orchestrated by LangGraph (StateGraph, parallel Bull + Bear via asyncio.gather). Schemas are Pydantic, Vertex-enforced on the Gemini side; Featherless responses pass through a markdown-fence stripper before parsing.
| Layer | Tech | Notes |
|---|---|---|
| Frontend | Next.js 16 · React 19 · Tailwind 4 · GSAP | App Router, Turbopack default. JavaScript (not TS) per author pref. |
| 3D scenes | Spline + @splinetool/react-spline | Two community scenes for hero accents (robot + interactive charts). |
| Agents | LangGraph 1.2 · Pydantic 2 · google-genai 2.3 · Featherless OpenAI-compat |
Module-scope Annotated[dict, reducer] state for parallel writes. |
| Inference | Gemini 2.5 Pro (Vertex AI, location="global") · Qwen3-32B (Featherless) |
1 M context, structured output via response_schema=<PydanticModel>. |
| Audio | yt-dlp · ffmpeg · Speechmatics (batch, diarized) | Word-level timestamps + speaker labels for click-to-jump UI. |
| Fundamentals | Financial Modeling Prep /stable/ |
Free tier; profile + ratios + 3 statements per ticker. |
| Filings | SEC EDGAR | 10-K + 10-Q stripped to plain text via lxml. |
| Host | Vultr High Frequency 8 vCPU · Ubuntu 24.04 · nginx · systemd | Frontend live; backend deploy planned post-agent build. |
| Layer | State |
|---|---|
| Ingestion pipeline (EDGAR + FMP + yt-dlp + Speechmatics) | ✅ Live; NVDA, TSLA, PLTR cached |
| Autonomous tier-scored audio candidate selector (T1 → T4) | ✅ Live; per-candidate audit in manifest.sources.audio |
Pydantic schemas (agents/schemas.py) |
✅ Citation, Claim, FilingAnalysis, CallAnalysis, BullCase, BearCase, Reconciliation |
Research questions (docs/RESEARCH.md) |
✅ All five resolved before agent code |
| Filing / Call / Bull / Bear / Reconciler agents | ✅ Shipped (LangGraph parallel-write reducer, per-node cache reuse) |
| FastAPI backend on Vultr (REST + SSE + ranged audio) | ✅ Live behind nginx + systemd |
Frontend (landing + /research/[ticker] dashboard + /methodology) |
✅ Live at https://diligence.duckdns.org |
| Adversarial security review (CORS, X-Forwarded-For, SSE dedupe, atomic manifest writes, SSRF defence) | ✅ 154 findings, top-severity all fixed |
| Day-5 UI revamp — 7 surgical sessions | ✅ Merged + deployed |
| HTTPS via DuckDNS + Let's Encrypt | ✅ https://diligence.duckdns.org, auto-renew |
| Demo video | 🟡 Pending — only blocker for hackathon submission |
| # | Stage | Shipped | Receipt |
|---|---|---|---|
| 1 | The Bet — primary-source cache | NVDA, TSLA, PLTR fully ingested. 364K-char 10-K, 170K-char 10-Q, ~60min earnings call, 10k+ word-level diarized transcript tokens per ticker. | data/{T}/{10k,10q,fundamentals}.json + earnings_call.mp3 + transcript.json + manifest.json |
| 2 | The Brains — five adversarial agents | Filing Analyst (~30 atomic claims, char-range citations). Call Analyst (~27 claims + hedging examples, speaker + timestamp citations). Bull + Bear (Featherless Qwen3-32B) argue from shared claim pool. Reconciler ranks disputed facts 1–10 by materiality. | analysis_{filing,call,bull,bear}.json + reconciliation.json. PLTR: 3 disputed facts at materiality 10/8/8, 100% transcript diarization. |
| 3 | The Cockpit — dashboard | Next.js 16 + React 19 + Tailwind 4 + Recharts + wavesurfer.js v7. Three-column bull / disputed / bear; click chart bar → swap focus card; click claim chip → seek transcript audio or open SEC URL. | Live: https://diligence.duckdns.org/research/PLTR |
| 4 | The Stress Test — adversarial review + autonomous audio selector | 3 Claude sub-agents ran in parallel dismantle mode: 154 findings (4 CRITICAL, 7 HIGH). CORS lockdown, X-Forwarded-For LAST-hop parsing, SSE seq dedupe, atomic manifest writes, SSRF defence on audio fetches, yt-dlp --dump-json. Autonomous YouTube candidate selector: 5-dim scoring across ≤16 candidates, tier T1 issuer-named → T4 unverified, gated by MIN_CANDIDATE_SCORE=50. |
services/audio.py:find_best_audio_candidate + manifest.sources.audio.candidates_considered |
| 5 | The Reveal — UI polish + HTTPS | 7 surgical UI sessions (per-claim ⚠/§ badges, ?fact=N permalink, claim chip → seek/SEC URL, dashboard-wide citation interactivity, audit candidates table polish). HTTPS via DuckDNS + Let's Encrypt + certbot auto-renew on Vultr nginx. |
https://diligence.duckdns.org (Let's Encrypt cert, expires 2026-08-15, auto-renew) |
| Service | Component | Cost (fresh ingest) | Notes |
|---|---|---|---|
| Vertex AI Gemini 2.5 Pro | Filing Analyst | ~$0.13 | 91K input + 5K output tokens |
| Vertex AI Gemini 2.5 Pro | Call Analyst | ~$0.05 | 25K + 3K |
| Vertex AI Gemini 2.5 Pro | Reconciler | ~$0.03 | 10K + 3K |
| Speechmatics | Diarized batch ASR (enhanced) | ~$1.04 | ~1 hr call × $1.04/hr tier-1 |
| Featherless | Bull + Bear (Qwen3-32B) | $0 marginal | Flat $25/mo plan |
| FMP / SEC EDGAR / yt-dlp | Fundamentals + filings + audio source | $0 | Free tier / public |
| Total | ~$1.25 |
Cached re-runs (agents.run TICKER --reuse-cache): $0 — every node short-circuits on data/{T}/analysis_*.json.
mirel-leonard-org enforces constraints/iam.disableServiceAccountKeyCreation, so service-account JSON keys cannot be created. The laptop's Application Default Credentials file is also owned by a separate production project (RobotBoy) — overwriting it would re-route that project's billing.
Solution: mint a short-lived OAuth access token via gcloud auth print-access-token against the active gcloud configuration. Token lives in-process, is auto-refreshed before expiry, never touches disk. ADC file is untouched.
Implementation: vertex_client.py. All Vertex AI work imports get_client() from that module.
gcloud config configurations create hackathon
gcloud config configurations activate hackathon
gcloud auth login your-email@example.com
gcloud config set project project-2be42b84-14e0-421a-b3a
gcloud services enable aiplatform.googleapis.com \
--project=project-2be42b84-14e0-421a-b3acd ~/diligence
source .venv/bin/activate
gcloud config configurations activate hackathongit clone https://github.com/leonardtudor11/diligence.git
cd diligence
# Backend
python3.12 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env # fill in API keys
# Smoke test all five services (Day-1 probes)
python scripts/probe_vertex.py
python scripts/probe_speechmatics.py
python scripts/probe_featherless.py
python scripts/probe_fmp.py
python scripts/probe_edgar.py
# End-to-end ingestion (NVDA cached)
python -m services.ingest NVDA
# Frontend
cd frontend
npm ci
npm run dev # → http://localhost:3000Copy .env.example → .env and fill in:
| Var | Source |
|---|---|
SPEECHMATICS_API_KEY |
speechmatics.com console (AIWEEK200 promo applied) |
FEATHERLESS_API_KEY |
featherless.ai dashboard |
FMP_API_KEY |
financialmodelingprep.com free tier |
SEC_USER_AGENT |
EDGAR ToS — set to your real email |
TWENTYFIRST_API_KEY |
optional — only if using 21st.dev Magic MCP for UI generation |
diligence/
├── vertex_client.py # Auth helper — gcloud token, never touches ADC
├── .env.example # Service credentials template
├── requirements.txt
├── scripts/ # Day-1 probes + research_probes.py (Day-2 RQ verification)
├── services/ # Production ingestion clients (Day 1 evening +)
│ ├── ingest.py, edgar.py, fmp.py, audio.py, speech.py
├── agents/ # Adversarial agents + LangGraph orchestrator (Day 2)
│ └── schemas.py # Pydantic contracts every agent reads/writes
├── frontend/ # Next.js 16 landing + dashboard (Day 3)
│ ├── app/components/ # Hero, BullBearSplit, SplineSceneDemo, Footer
│ ├── components/ui/ # Card, Spotlight, Spline scene wrapper
│ ├── lib/utils.js # cn() helper
│ └── public/ # Bull/bear silhouettes + ticker brand logos
└── docs/
├── CONTEXT.md # Original project briefing
├── HANDOFF.md (../) # Per-session handoff doc
├── AUDIT.md # Day-1 adversarial review findings
├── RESEARCH.md # Day-2 research-question answers
└── screenshot-landing.png # Hero screenshot for this README
- Async Gemini surface:
await client.aio.models.generate_content(...)works — no thread wrapping needed. - Structured output:
response_schema=<PydanticModel>enforced by Vertex for deeply nested lists. - Featherless gotcha:
response_format: {"type": "json_object"}is accepted but Qwen3-32B wraps payload injson.loads. - LangGraph parallel writes: module-scope
TypedDictwithAnnotated[dict, _merge_agents](Python 3.14 strict forward-ref eval). - Prompt-injection defence: wrap filing/transcript content in
<filing>/<transcript>XML, system prompt treats tag contents as data not instructions, surfaceinjection_detected: boolon every analyst output.
| Track | What we use |
|---|---|
| Vultr | High Frequency 8 vCPU/16 GB instance in Frankfurt, $200 credit covers ~2 months runtime |
| Gemini | gemini-2.5-pro for Filing / Call / Reconciler (heavy reasoning + structured output) |
| Featherless | Qwen3-32B for parallel Bull / Bear adversarial reasoning |
| Speechmatics | Batch transcription + speaker diarization on the earnings call MP3 |
.envis gitignored; the example file ships only var names + free-tier acquisition instructions.- Auth uses short-lived gcloud OAuth tokens — no SA keys, no ADC file mutation.
httpxandhttpcoreloggers muted in every module that hits a query-string API to prevent secret leakage in tracebacks.- All HTTP-response key access wrapped + raised as a domain-specific
DiligenceErrorsubclass (audit-fixed Day 1). - Production VM: UFW restricts to 22 / 80 / 443, fail2ban on sshd, automatic unattended security updates.
Full policy: SECURITY.md. Day-1 audit report: docs/AUDIT.md.
MIT — see LICENSE.
Third-party visual assets retain their original licenses; see frontend/CREDITS.md (game-icons CC BY 3.0, Simple Icons CC0, Spline community CC0).
- lablab.ai / Milan AI Week '26 for the hackathon track + sponsor credits.
- Google Vertex AI, Featherless, Speechmatics, Vultr for the model + audio + compute credits.
- Lorc for the bull and bear silhouettes (game-icons.net, CC BY 3.0).
- 21st.dev for the Spline-component pattern that ships our 3D hero accents.
🤖 Built with Claude Code (Anthropic), pair-programming in the loop.