Federal Opportunity Intelligence Platform
A 27-tool cognitive architecture built directly on Claude's API with constitutional AI governance, self-improving evaluation loops, and structural hallucination prevention. Automated opportunity scoring, workforce matching, and pursuit pipeline management for government contractors.
Opportunity Discovery
Automated federal opportunity ingestion and enrichment
- Automated Federal Ingestion — Pulls thousands of federal solicitations from SAM.gov daily, filtering irrelevant opportunities
- AI Document Enrichment — Reads RFPs, SOWs, and amendments to extract structured data including scope, staffing, compliance, and evaluation criteria
- Data Normalization — Resolves messy federal data into a canonical dataset of locations, company names, and NAICS codes
- Award & Contract Tracking — Integrates USASpending to track contract awards and link them back to original solicitations
- Multi-Tier Scoring — Three layers of relevance scoring at ingestion, office-level, and capability card levels
Strategic Decisioning
AI-driven intelligence for pursuit decisions
- Predictive Intelligence — 11 prediction types including win probability, competitive field size, deadline slip, and protest likelihood
- GAO Protest Intelligence — Vector database of GAO protest decisions with retrieval-augmented search for precedents and risk identification
- Competitive Landscape — 2,000+ competitor profiles from USASpending, hiring intelligence, and entity graph mapping 122,000+ relationships
- Sentinel AI Assistant — Dedicated AI agent per opportunity with 16 specialized tools and natural conversation interface
- Intelligence Briefs — 13-section AI-generated capture document synthesizing enrichment, competitive landscape, predictions, and staffing
Capture Execution
Workforce matching, staffing, and pursuit management
- Workforce Matching — Scores entire workforce against opportunity staffing requirements including clearances, certifications, skills, and availability
- Adaptive AI Interviews — AI discovers hidden qualifications not in HR records, feeding findings back to improve match scores
- Bench Visibility & Utilization — Real-time workforce availability tracking, surfacing idle capacity for placement
- Pipeline & Pursuit Management — Kanban-style pipeline from identification through award with stage gates and action items
- HR Integration — Bidirectional connection to HR systems for workforce data synchronization
Agent Architecture
Built directly on Anthropic's Claude API with no orchestration frameworks — owning the full stack from persona definition to tool dispatch to evaluation.
Constitutional AI Governance
900-line constitutional document (soul.yaml) as a single source of truth: persona definition, confidence calibration rules that force language certainty to match data quality, strategic decision frameworks, and a zero-tolerance banned-word list. Behavior changes ship as YAML diffs, not model changes.
Self-Improving Evaluation
After-Action Review agent evaluates every session across 6 diagnostic dimensions with tier-appropriate rubrics. An optimizer uses Opus to propose mutations to tool definitions, constitutional rules, and guardrails, A/B testing each variant against a 20-metric scoring rubric with safety gates.
Structural Hallucination Prevention
Mandatory data gap disclosure forces '[DATA GAP]' statements when tools return empty. CRAG grades retrieval relevance and re-retrieves on low-confidence results. Diminishing-returns detection stops search spirals after consecutive rounds with no new information.
Tool Disambiguation & Routing
27 tools with explicit USE WHEN / DO NOT USE disambiguation rules, recursion prevention, and cascade hierarchies that funnel broad queries to the correct specialized tool. Tier-aware model routing (Haiku/Sonnet/Opus) via zero-cost regex classification with prompt caching.
66-Table Data Platform
PostgreSQL schema covering opportunities, competitors, employees, predictions, intelligence briefs, and audit trails. Integrated with USASpending for award tracking and entity graph mapping 122,000+ relationships across the competitive landscape.
Goodhart's Law Defense
Each evaluation layer grounds the layer below it against progressively more fundamental constraints, so no single optimization loop defines its own success criteria. The system can propose changes to its own scoring metrics, but these require human approval.