Ramps Logistics / Procurement Intelligence
A living procurement intelligence system built for Ramps Logistics. A swarm of specialized AI agents filters 50,000 SAM.gov opportunities, sources suppliers, drafts bids, and learns from every decision you make. It gets smarter with every run.
Current State
The team uses a custom scraper to search SAM.gov. It pulls ~50,000 opportunities and narrows them to ~935 using NAICS/PSC codes and a fit scoring algorithm.
For promising opportunities, the team reads through the RFQ details (often 200+ pages) looking for critical information: submission platform, country of origin requirements, compliance rules, and exact specs.
Once an RFQ looks viable, the team downloads the PDF again and uploads it to a custom GPT tool that extracts the items needed and searches for suppliers.
The team manually drafts emails to potential suppliers asking for pricing confirmation, availability, and lead times.
The team analyzes the application requirements (which are different for every RFQ) and manually drafts the bid response.
The team reviews and submits the bid.
Gap Analysis
The SamWise Pipeline
The Sub-Agent Swarm
Each agent is purpose-built for a single task. They work in parallel, share context, and improve independently. Upgrade one without touching the others.
This is not about replacing your team. It is about removing the monotonous, manual work so your people become strategic decision-makers instead of document readers and email writers.
Scans ~50,000 SAM.gov opportunities and returns only the ones Ramps can and should bid on.
When the team skips an opportunity the agent scored highly, or pursues one scored low, the scoring model adjusts. False positives and false negatives feed back into the filter rules.
Monitors opportunities the team is actively working on. Sends immediate alerts when anything changes.
Tracks which types of changes matter most to the team (e.g., spec changes vs. deadline extensions) and prioritizes alerts accordingly.
Takes an RFQ, extracts every item that needs to be procured, finds suppliers, and ranks them using a weighted scoring model.
Every sourcing run adds new supplier data. After-action reports on supplier performance feed back into scoring. Over time, the agent knows which suppliers deliver on time, at the best total cost, and recommends them first.
Drafts professional inquiry emails to potential suppliers.
Tracks which suppliers respond, how quickly, and how competitive their pricing is. This data feeds back into the sourcing agent's supplier ranking.
Analyzes the application requirements for a specific RFQ and drafts the bid response.
As bids are submitted, win/loss outcomes are tracked. The agent identifies patterns in successful bids and adjusts its drafting approach.
Architecture
Most procurement tools use hard-coded rules or disconnected GPT sessions. SamWise uses a coordinated swarm of AI agents that share context, learn from outcomes, and improve independently.
The Filter Agent passes scored opportunities to the Watch Agent and the Sourcing Agent. The Sourcing Agent downloads documents, extracts items, checks the supplier database, and passes results to the Outreach Agent. The Outreach Agent drafts supplier emails and tracks responses. Once suppliers confirm, the Bid Draft Agent pulls everything together: RFQ requirements, sourcing data, supplier confirmations, and Ramps' company profile. Each agent reads from and writes to a shared context layer, so no information is lost between steps. The system remembers every interaction, every supplier response, every bid outcome.
Disqualifier Engine
Instant. No document download needed.
| Rule | What It Catches |
|---|---|
| Ineligible set-aside | Opportunities with set-aside requirements that Ramps does not qualify for |
| Deadline too short | Bids due in 2 days or less |
| Excluded PSC categories | Weapons, ammunition, aerospace, vehicles, IT/telecom, agricultural, construction, medical, R&D, education, and 20+ other categories outside Ramps' scope |
| Inactive notices | Past awards, market research, sources sought (not actual bidding opportunities) |
AI reads the attached documents and flags these.
| Rule | What It Catches |
|---|---|
| DIBBS submission required | Opportunities that must be submitted through DLA's DIBBS platform |
| Complex/combo scope | Purchase + deliver + train + install bundles or service-based scopes |
| Non-USA location | Small business set-asides for non-US places of performance |
| Requires US-based assets | RFQs requiring ownership of US-based equipment, vehicles, warehouses, or facilities that Ramps does not have |
Output: Each opportunity gets a pass/fail verdict. If disqualified, the exact rule and the quoted text that triggered it are shown.
The Living System
The team reviews the agent's shortlist. Opportunities they skip despite high scores, or pursue despite low scores, create feedback signals. The scoring model adjusts. Over time, the filter gets more accurate at predicting which RFQs the team will actually bid on.
Every sourcing run adds new data: supplier, pricing, lead time, and reliability score. After-action reports capture on-time delivery rates and total landed costs. After 50 runs, the agent ranks suppliers by weighted performance. After 200 runs, Ramps has a proprietary supplier intelligence database that no competitor can replicate.
As supplier responses flow back, the system tracks response rates, pricing competitiveness, and reliability. The sourcing agent uses this data to rank suppliers: "For category X, Supplier A responds in 24 hours at 15% below market. Start there."
Win/loss outcomes train the bid drafter. Successful bids reveal what evaluators prioritize. Failed bids reveal what to avoid. After enough data, the agent can predict win probability for a given opportunity before the team invests time.
After 6 months of operation, SamWise will have:
Evolution
Each phase builds on the last. The system earns more autonomy as it proves itself. Phase 2 is not a separate product. It is Phase 1 after enough learning loops have run.
The key insight: Phase 2 is not something we build separately. It is what Phase 1 becomes after enough data flows through the learning loops. Every decision the team makes in Phase 1 trains the system for Phase 2. The path to autonomy is built into the architecture from day one.
Impact
| Metric | Current | Target (90 Days) | Improvement |
|---|---|---|---|
| Time per RFQ cycle | 4-8 hours manual | Under 30 minutes | ~90% reduction |
| Disqualified opportunities caught | ~60% (many slip through) | 95%+ | +35% coverage |
| Opportunities reviewed per day | 10-20 | 100+ | 5-10x throughput |
| Sourcing workflow | Manual download + upload to GPT | One-click automated | Fully integrated |
| Supplier search time | 1-2 hours per RFQ | Under 5 minutes | ~95% reduction |
| Bid draft time | 2-4 hours per RFQ | Under 15 minutes | ~90% reduction |
The Vision
SamWise surfaces the best opportunities, ranks them by likelihood of success based on past bid data, and presents a complete activation plan. Your team approves. The system executes.
Conceptual dashboard showing the ideal state. Win probabilities are calculated from historical bid data, supplier confirmations, and competitive analysis.
Next Steps
For direct API access to opportunity data
PDFs with known sourcing outcomes for eval testing
Which certifications Ramps currently holds
Currently used in the existing scraper
For watch alerts: email, Slack, or other
Five items above are all we need. Once provided, the first agent (Filter) can be live within two weeks.