BidSmith — Federal RFP Intelligence Platform

by ARIS Labs · bidsmith.pro

AI-powered audit engine for government contractors. Instant compliance matrix, FAR/DFARS risk flags, and bid/no-bid recommendation from any SAM.gov link or PDF.

---

Quick Start

Golden Run — one command to start everything

npm run dev & npm start

• npm run dev — Vite dev server on :5173 (hot-reload, proxies /api → :8080)
• npm start — Express API server on :8080

Note: This is a Vite + React frontend with an Express.js API backend. There is no Python/uvicorn server. The API lives in api/index.js.

Environment

Copy .env.example to .env and fill in:

OPENROUTER_API_KEY=         # LLM inference
VITE_CLERK_PUBLISHABLE_KEY= # Auth
CLERK_SECRET_KEY=
SAM_GOV_API_KEY=            # Optional — SAM.gov enrichment

---

Bento Dashboard

The intelligence dashboard lives at /bento. It provides:

Card	Route wired
RFP Upload Zone	POST /api/audit/pdf · POST /api/audit/link
Live Analysis (confidence scores + citations)	Reactive to upload result
Bid Brief + Copy to Clipboard	Reactive to upload result
Eval Status (5 dots, live poll)	GET /api/evals/status

---

Inference Regression Testing (Evals)

Run the Golden Set against the live API before shipping any prompt or RAG change.

Run evals (local)

python scripts/run_evals.py

Run against production

EVAL_API_URL=https://bidsmith.pro python scripts/run_evals.py

Save a baseline (do this after a known-good state)

python scripts/run_evals.py --baseline

Compare against baseline (regression check)

python scripts/run_evals.py --compare --verbose

CI gate (exits 1 if overall fails)

python scripts/run_evals.py || echo "REGRESSION DETECTED — do not ship"

Output

Results are written to tests/eval_results.json and immediately reflected in the /bento dashboard Eval Status card (polls every 30 seconds).

The 5 Golden Evals

#	Label	Tests
1	SHALL vs WILL Compliance Extraction	Parsing precision — must distinguish SHALL, WILL, ambiguous
2	Bid/No-Bid Under Competing Constraints	Reasoning under conflicting eligibility signals
3	Cross-Section Requirement Conflict	Cross-document reasoning, FAR section hierarchy
4	NIST SP 800-53 Scope Interpretation	Domain knowledge depth (cybersecurity controls)
5	Teaming Agreement Compliance	FAR 52.219 limitation on subcontracting, affiliation risk

Rule: If any eval regresses below its min_keyword_score, do not ship.

---

API Endpoints

Method	Path	Description
POST	/api/audit/pdf	Upload PDF → full audit result
POST	/api/audit/link	SAM.gov URL → full audit result
POST	/api/audit/text	Raw text → full audit result (used by evals)
GET	/api/evals/status	Serves tests/eval_results.json to the dashboard
GET	/api/health	API + provider key status
POST	/api/chat	Follow-up chat with audit context

Audit Pipeline — 3 Logged Stages

Every /api/audit/pdf and /api/audit/link call emits structured JSON logs:

stage: 1_retrieval  — PDF parse or SAM.gov fetch
stage: 2_chunking   — Text preparation and size estimation
stage: 3_inference  — 3-agent LLM pipeline (Extractor → Auditor → Strategist)

Search logs for "audit_stage" to trace any request end-to-end.

Error format

All API errors return consistent JSON:

{ "error": "Human-readable message", "code": 422 }

---

Project Structure

aris-core/
├── api/                    # Express.js API
│   ├── agents/             # 3-agent pipeline (coordinator, extractor, auditor, strategist)
│   ├── services/           # SAM.gov, analytics, waitlist, stripe
│   ├── utils/              # logger, tracing, mailer, asyncHandler
│   └── index.js            # Entry point, all routes
├── src/                    # Vite + React frontend
│   ├── components/
│   │   └── bento/          # Bento dashboard components
│   │       ├── RfpUploadZone.jsx
│   │       ├── LiveAnalysisCard.jsx
│   │       ├── BidOutputCard.jsx
│   │       └── EvalStatusCard.jsx
│   └── pages/
│       ├── BentoDashboard.jsx   # /bento
│       └── GovConDashboardV2.jsx # /app (full chat UI)
├── scripts/
│   └── run_evals.py        # Inference regression tester
└── tests/
    ├── evals.json           # Golden Set definitions
    └── eval_results.json    # Latest run output (auto-generated)

---

Deployment

Deployed on Railway (API + static). Primary CI is .github/workflows/ci.yml.

# Build
npm run build

# Preview build locally
npm run preview

---

Force Railway rebuild — see railway.json for config.

---

Aris registry & Python SDK (optional)

This repository also ships the aris-sdk Python package: a small registry + worker-node stack for decentralized LLM sessions (ai.generate, ai.chat), balance/usage APIs, and JWT handshakes. It does not replace the Node API above; it lives alongside it for experimentation and PyPI distribution.

Install (library)

pip install aris-sdk

Quick example

from aris.client import Aris

client = Aris(api_key="aris_live_...")
print(client.generate("Why is the ocean salty?"))
print(client.chat([{"role": "user", "content": "Capital of France?"}])["content"])

Run registry + worker locally

See .env.example (section Aris registry / Python SDK) and CONTRIBUTING.md. Pytest targets aris/tests/ and runs in .github/workflows/aris-python-ci.yml when Python paths change.

Published package metadata still uses this README; keep both sections accurate.