AI-powered procurement intelligence for the City of Richmond.
Built for Hack for RVA 2026 | Pillar 1: A Thriving City Hall | Problem 2: Procurement Risk & Opportunity Review
Live Demo: hackrva.ithena.app | GitHub: team-aether
City of Richmond procurement staff manage 1,365 contracts worth $6.76 billion across 37 departments. Today, reviewing a single contract for renewal requires:
- Manually searching multiple databases (City, State VITA, Federal SAM.gov)
- Reading through hundreds of pages of PDF contract documents
- Checking federal compliance lists (FCC, DHS, FBI, FTC) one by one
- Comparing vendor pricing with no centralized tool
- Tracking expiration dates across spreadsheets and emails
Result: Missed renewals, expired contracts still in use, no visibility into vendor concentration risk, and millions in potential savings left on the table.
RVA Contract Lens is a Federated Decision Intelligence platform — 8 independent data sources aggregated into a single decision interface with full AI transparency. It transforms a 60-minute manual review into an 8-second decision brief that shows exactly what data the AI analyzed, how each source influenced the verdict, and why.
One platform, two audiences.
Staff see what needs action — not a data dump, but three urgency lanes: Decide Today, Plan This Week, Review This Month. Every card links directly to the AI Decision Engine.
Not a dashboard feature — an architectural pattern. The Decision Engine federates 8 independent intelligence sources into a single decision interface. Each source is queried in parallel, weighted by relevance, and synthesized through a structured LLM call that produces a verdict with full evidentiary transparency.
The pattern is domain-agnostic: swap the data sources and the same architecture works for grants management, permitting, fleet maintenance, or any civic domain where staff make recurring decisions across fragmented systems.
For procurement, select a vendor and contract. The system runs all 8 intelligence sources in parallel — you can watch each one complete in real time:
Then delivers a RENEW / REBID / ESCALATE verdict — with every data point cited:
8 federated intelligence sources:
| Source | Type |
|---|---|
| Contract details | DuckDB (value, dates, department, risk level) |
| Vendor history | All contracts for this vendor across departments |
| SAM.gov compliance | Live federal API — exclusions and opportunities |
| FCC Covered List | Prohibited manufacturers (Huawei, ZTE, etc.) |
| Consolidated Screening List | DHS, FBI, FTC sanctioned entities |
| Vendor concentration risk | HHI index — monopoly risk by department |
| PDF contract terms | OCR-extracted clauses via semantic search (ChromaDB) |
| Vendor web intelligence | DuckDuckGo public reviews, news, reputation |
Three-layer output (progressive disclosure):
- Traffic light — RENEW / REBID / ESCALATE with confidence score (quick glance)
- Evidence grid — pros and cons with cited data sources, signed impact factors (+20 Compliance Clear, -15 Price Increasing), equity & MBE context (full analysis)
- Decision memo — exportable document with all evidence, alternative vendors, and recommended next steps (institutional record)
The AI recommends. Humans decide. Every verdict is persisted to an audit trail — building institutional memory so the next procurement officer inherits data-driven context, not a blank slate.
Grades all 37 departments A through F based on contract risk. Collapsible sections for expiry forecast, risk distribution, top anomalies, and department grades.
Models the fiscal impact of rebidding concentrated contracts. Three scenarios (5% / 10% / 15%) with specific vendors to target, projected savings per department, and actionable buttons that link directly to the Decision Engine.
AI-generated procurement strategy per department: how many contracts to renew, rebid, or escalate — with projected savings, equity context, and specific action items with links.
Every AI decision is recorded. Builds institutional memory so the next procurement officer inherits data-driven context, not a blank slate. Pattern analysis surfaces trends across the portfolio.
Herfindahl-Hirschman Index analysis identifies departments over-dependent on single vendors. Flags monopoly risk before it becomes a crisis.
Tracks vendor diversity ratios by department, small business participation, and competitive bidding rates. Equity context is embedded in every Decision Engine analysis — not isolated on a separate page.
Upload a scanned contract PDF. OCR extracts the full text (176K+ chars from a single document), then AI identifies key terms: value, expiration, renewal clauses, and parties.
Type a question in plain English: "Show me expiring IT contracts over $100K." AI translates it to a database query and returns results instantly.
Risk matrix maps anomalies by severity and likelihood. Each anomaly includes remediation steps and links to the relevant analysis tool.
Searchable, sortable, filterable table of all 1,365 contracts. Mobile-responsive card layout. Click any contract to analyze in the Decision Engine.
Residents can explore $6.76 billion in city spending by department, vendor, or service — without filing a FOIA request.
Interactive charts showing spending by department. Click any department or vendor to drill into their full contract portfolio.
Additional public pages: Vendor directory, department detail, Find a Service (20+ categories), and Data Sources (transparency trust signal).
RVA Contract Lens is designed to work for everyone, not just the easiest users.
Every page includes a persistent accessibility toolbar — font scaling, high contrast, bilingual toggle, and dark mode — available to both staff and the public:
- Bilingual EN/ES across all staff pages (30+ translated strings)
- Atkinson Hyperlegible font — designed for low-vision and dyslexia accessibility
- Font size toggle (A, A+, A++) and high-contrast mode on both staff and public pages
- Skip links on every page (bilingual) for keyboard and screen reader users
- 44px minimum touch targets throughout, with
prefers-reduced-motionsupport - Dark mode for reduced eye strain and low-light environments
- MBE/equity context embedded in the AI Decision Engine — not a separate page, but part of every analysis
- Public transparency view so residents with any level of digital literacy can explore city spending
| Metric | Value |
|---|---|
| Real contracts analyzed | 1,365 ($6.76B) |
| Data sources in Decision Engine | 8 |
| Federal compliance checks | 3 automated (SAM.gov, FCC, CSL) |
| Staff pages | 23 |
| Public transparency pages | 7 |
| API endpoints | 46 |
| Backend routers | 12 |
| Commits | 172+ |
| Team size | 4 members, 36 hours |
| Manual review time saved | 60 min → 8 sec per contract |
| OCR extraction capacity | 176K+ chars from a single scanned PDF |
| Source | Records | Method | Update Frequency |
|---|---|---|---|
| City of Richmond (Socrata) | 1,365 contracts ($6.76B) | CSV download | Real-time available |
| SAM.gov Federal | Live query | Opportunities API | Real-time |
| eVA Virginia State | Seeded | State procurement data | Periodic |
| VITA IT Contracts | Seeded | State IT procurement | Periodic |
| FCC Covered List | 10+ entities | Keyword matching | Updated with FCC releases |
| Consolidated Screening List | 10+ entities | Keyword matching | Updated with federal releases |
| Hackathon PDFs | 10 contracts (206 chunks) | OCR via unstructured |
On upload |
| DuckDuckGo Web Intel | Live search | HTML scraping | Real-time |
# Clone
git clone https://github.com/ankitSrivastavaITH/team-aether.git
cd team-aether
# Backend
cd backend
pip install -r requirements.txt
echo "GROQ_API_KEY=your_key_here" > .env # Get free key at console.groq.com
python -m uvicorn app.main:app --port 8200 --reload
# Frontend (new terminal)
cd frontend
npm install
npm run dev -- -p 3200- Frontend: http://localhost:3200
- Backend API docs: http://localhost:8200/docs
- Get a free Groq API key at console.groq.com
┌─────────────────────────────────────────────────────────────────────────┐
│ DATA INGESTION LAYER │
├──────────────┬──────────────┬──────────────┬──────────────┬────────────┤
│ Socrata │ SAM.gov │ eVA / VITA │ PDF Upload │ Web Scrape│
│ CSV API │Opportunities │ State CSV │ (contract │ DuckDuckGo │
│ 1,365 rows │ Live API │ Seeded │ documents) │ Lite │
│ ($6.76B) │ │ │ │ │
└──────┬───────┴──────┬───────┴──────┬───────┴──────┬───────┴─────┬──────┘
│ │ │ │ │
v v v v v
┌─────────────────────────────────────────────────────────────────────────┐
│ PROCESSING LAYER │
│ │
│ ┌──────────────────┐ ┌──────────────────┐ ┌──────────────────────┐ │
│ │ CSV → DuckDB │ │ OCR Pipeline │ │ Compliance Engine │ │
│ │ Ingest scripts │ │ `unstructured` │ │ │ │
│ │ (pandas/polars) │ │ library │ │ SAM.gov API (live) │ │
│ │ │ │ 176K+ chars │ │ FCC Covered List │ │
│ │ city_contracts │ │ per document │ │ Consolidated Screen │ │
│ │ state_contracts │ │ │ │ │ List (DHS/FBI/FTC) │ │
│ │ vita_contracts │ │ v │ │ │ │
│ └────────┬─────────┘ │ Chunking + │ └──────────┬───────────┘ │
│ │ │ Embedding │ │ │
│ │ └────────┬──────────┘ │ │
└───────────┼─────────────────────┼────────────────────────┼──────────────┘
│ │ │
v v v
┌─────────────────────────────────────────────────────────────────────────┐
│ STORAGE LAYER │
│ │
│ ┌──────────────────────────┐ ┌──────────────────────────────────┐ │
│ │ DuckDB │ │ ChromaDB │ │
│ │ (Embedded Analytics) │ │ (Vector Store) │ │
│ │ │ │ │ │
│ │ city_contracts (1,365) │ │ pdf_chunks (206 embeddings) │ │
│ │ state_contracts │ │ semantic similarity search │ │
│ │ vita_contracts │ │ used by Decision Engine for │ │
│ │ extracted_terms (9) │ │ contract term cross-reference │ │
│ │ decisions (audit trail) │ │ │ │
│ └──────────────────────────┘ └──────────────────────────────────┘ │
└───────────────────────────┬────────────────────────────────────────────┘
│
v
┌─────────────────────────────────────────────────────────────────────────┐
│ FastAPI BACKEND (12 Routers) │
│ │
│ /contracts /decision /strategy /health-scan /analytics │
│ /nl-query /extract /insights /mbe /services │
│ /parser │
│ │
│ Key: async parallel compliance checks, rate limiting (slowapi), │
│ decision persistence, robust JSON fallback parsing │
└───────────────────────────┬────────────────────────────────────────────┘
│
v
┌─────────────────────────────────────────────────────────────────────────┐
│ AI / LLM LAYER (Groq) │
│ │
│ Model: llama-3.3-70b-versatile (primary) │
│ Fallback: llama-3.1-8b-instant (rate limit cascade) │
│ Recovery: Regex extraction from malformed JSON │
│ │
│ Used for: │
│ - Decision Engine (8-source → RENEW/REBID/ESCALATE verdict) │
│ - NL-to-SQL translation (plain English → DuckDB query) │
│ - Contract term extraction (PDF → structured fields) │
│ - Risk summaries (portfolio-level AI insights) │
└───────────────────────────┬────────────────────────────────────────────┘
│
v
┌─────────────────────────────────────────────────────────────────────────┐
│ FRONTEND (Next.js 14) │
│ │
│ STAFF (23 pages) PUBLIC (7 pages) │
│ Dashboard, Decision Engine, Overview, Spending, │
│ Health Scanner, What-If, Vendors, Vendor Detail, │
│ Portfolio Strategy, Audit, Department Detail, │
│ Risk, MBE, PDF, Ask Richmond, Find a Service, Sources │
│ Contracts, Analytics, + more │
│ │
│ UI: shadcn/ui + Tailwind + Recharts + TanStack Query/Table │
│ a11y: Atkinson Hyperlegible, i18n EN/ES, skip links, 44px targets │
└───────────────────────────┬────────────────────────────────────────────┘
│
v
┌─────────────────────────────────────────────────────────────────────────┐
│ DEPLOYMENT: Docker Compose → Cloudflare Tunnel │
│ hackrva.ithena.app (frontend) + api-hackrva.ithena.app (backend) │
└─────────────────────────────────────────────────────────────────────────┘
| Layer | Technology | Why |
|---|---|---|
| Frontend | Next.js 14, React 18, Tailwind CSS, shadcn/ui, TanStack Query/Table, Recharts | Fast SSR, component library, real-time data fetching |
| Backend | FastAPI (Python 3.11), 12 routers | Async support, auto-generated API docs, fast development |
| Analytics DB | DuckDB (embedded) | In-process OLAP — no database server needed, handles 1M+ rows |
| Vector DB | ChromaDB | Semantic search over OCR'd PDF chunks |
| AI/LLM | Groq (llama-3.3-70b-versatile) | Free tier, fast inference (< 2 sec), structured JSON output |
| OCR | unstructured library |
Handles scanned PDFs, images, and mixed-format documents |
| Deployment | Docker Compose + Cloudflare Tunnel | Zero-infrastructure demo deployment |
| Decision | Rationale |
|---|---|
| DuckDB over PostgreSQL | Embedded analytics — zero ops, sub-second queries on 1,365 contracts, no server to manage |
| Groq over OpenAI | Free tier (500K tokens/day), fast inference, good JSON output from llama-3.3-70b |
| ChromaDB for PDF search | Lightweight vector DB, persistent storage, semantic similarity on OCR chunks |
| FastAPI over Flask | Async support critical for parallel compliance checks (SAM.gov + FCC + CSL concurrent) |
| shadcn/ui over Material | Accessible by default, composable, Tailwind-native, small bundle size |
| DuckDuckGo Lite over Google | Google blocks HTML scraping; DDG Lite returns parseable HTML results |
| Limitation | Impact | Mitigation |
|---|---|---|
| Groq free tier rate limits | AI Decision Engine may timeout under heavy concurrent use | 5-second pacing between calls during demo; upgrade to paid tier for production |
| SAM.gov Exclusions API requires Entity Management role | Cannot check debarment directly | Using Opportunities API as proxy; apply for Entity Management access for production |
| Trade.gov CSL API retired | Consolidated Screening List check uses offline keyword matching | Monitor for API replacement; current keyword list covers primary entities |
| DuckDuckGo web search ~3-5 seconds | Adds latency to Decision Engine analysis | Results cached; runs in parallel with other checks |
| MBE status not in public contract data | Cannot definitively identify MBE-certified vendors | Analyze diversity ratios and small business participation as proxy indicators |
| PDF OCR accuracy varies | Scanned documents with poor quality may extract incomplete text | Uses unstructured with multiple parsing strategies; manual review always available |
If piloted by the City, the following enhancements would maximize impact:
| Phase | Feature | Value |
|---|---|---|
| Pilot (0-3 months) | Connect to City's internal contract management system | Eliminate manual CSV imports, real-time data |
| Add SAM.gov Entity Management API access | Full debarment checking, not just opportunities | |
| Expand PDF corpus | Ingest all active contract PDFs for comprehensive term extraction | |
| Scale (3-6 months) | Multi-city deployment | Same platform, different Socrata dataset ID — works for any city with open data |
| Email/calendar integration | Auto-alert procurement officers 90/60/30 days before expiration | |
| Vendor self-service portal | Vendors update their own MBE certifications, contact info, and capabilities | |
| Transform (6-12 months) | Predictive risk scoring | ML model trained on historical contract outcomes to predict which contracts will have problems |
| Cross-jurisdiction price benchmarking | Compare Richmond's contract prices against other Virginia cities and state rates | |
| Automated RFP generation | AI drafts request-for-proposal documents from contract requirements | |
| Legislative compliance tracking | Monitor Virginia procurement law changes and flag affected contracts |
cd backend && pytest tests/ -v5 critical path tests covering health check, contracts API, compliance endpoints, health scanner, and NL query.
Built by Team Aether (4 members) for Hack for RVA 2026.
AI-generated recommendations are advisory only. All procurement decisions require human review and approval per City of Richmond procurement policy.