Selected Work

Production systems shipped for defense contractors, industrial holding companies, and proprietary trading and SaaS infrastructure. Each engagement is forward-deployed — I build directly in your codebase, on your timeline.

Featured Platform Reputation Surveillance Platform

Counter-Narrative Intelligence

Counter-Narrative Intelligence Platform

Visit argus.forwardai.dev →
OUTCOME FOR THE ENTERPRISE UNDER ATTACK Defend the company — without becoming the disinformation. Every counter-narrative ships with: PV-GRADE FORMAT 21 CFR 201.57 canonical Label 16 sections, GRADE. CIOMS frequencies. QPNV sign-off block. VALIDITY ENVELOPE 4 permits, 4 forbids Rank, flag, surface, grade. RefusalLabel triad. No efficacy claims. TWO-PLANE DATA Customer isolated, Commons shared AWS-account boundary. DPA-layered. Network-effect moat. FOUNDRY-COMPANION JSON-LD canonical, STIX 2.1 emission FMI-CK open spec. DISARM-isomorphic. Bitemporal replay.
What It Does
  • Ships four defensible artifacts per attack — Counter-Narrative Label, PSUR-N adverse-event report, Narrative Resilience Score, and Deconfliction Sheet
  • Gives Legal, Risk, and D&O insurance a sign-able artifact before any public counter-message goes out
  • Scores narrative resilience FICO-style (300–850) on legal persons only, with FCRA-style reason codes
  • Pre-tests counter-messages against a stochastic-completion harness before the company speaks
  • Keeps regulated customer data inside a single-tenant Customer Plane while the cross-customer Commons Plane builds the data-network-effect moat
Under the Hood
  • Next.js 15 (App Router + RSC) + TypeScript 5.7 strict + Prisma 5 over a 16-table multi-tenant schema
  • FMI-CK v0.1 ontology: DISARM-Red-isomorphic Tier A behaviors + Tier B doctrine fields under DISARM Foundation + Carnegie PCIO governance, CC-BY 4.0
  • Validity-envelope grammar: exactly 4 permitted claim shapes (rank, flag, surface, grade) and 4 forbidden (population effect, individual prediction, behavior change, replacement)
  • RefusalLabel / ScopedDowngrade / HumanPanelEscalation triad on every out-of-envelope request
  • Bitemporal-ready schema (event_id, valid_from, valid_to, recorded_at) with `/v1/replay/<deliverable_id>` for 21 CFR Part 11 byte-identical regeneration
  • One Pydantic schema co-emits `/v1/intelligence-cards` (JSON-LD, Foundry Ontology import) and `/taxii2/collections/fmi-ck-canonical` (STIX 2.1)
  • Modular monolith on Postgres + S3 + Airflow; Phase-0-boring infra, demo-grade Sprint 1 against the Permian Basin "fracking causes cancer" attack scenario
Enterprise AI Governance

AI Control Tower

A sovereign-grade governance backbone for the AI agents running inside the enterprise

OUTCOME FOR THE ENTERPRISE Govern every AI agent — without slowing teams down. Every enterprise AI program gets: VISIBILITY Every agent inventoried Zero shadow AI. Real-time drift alerts. Owners + consumers. CONTAINMENT Stage-disable, not panic-stop Circuit breakers. Graceful shutdown. Operators in control. CFO-GRADE ROI Board-ready ROI evidence Owner / consumer split. Defensible spend. Live program metrics. SOVEREIGNTY Regulated data stays inside GDPR / PIPL safe. Non-collection by design. Sovereign perimeter.
What It Does
  • Inventories every AI agent reaching into production systems
  • Contains shadow AI without slowing teams down
  • Gives the CFO defensible, board-ready ROI evidence for the AI program
  • Lets operators stage-disable agents instead of hitting an emergency stop
  • Keeps regulated data inside the sovereign perimeter (GDPR / PIPL)
Under the Hood
  • FastAPI core for agent registration, versioned passports (Agent : Version : Environment), and gateway telemetry
  • React dashboard for live metrics, drift alerts, and owner-vs-consumer chargeback
  • Python "Flight Control" SDK: circuit breaker, telemetry manager, graceful shutdown
  • Azure APIM policies: passport validation, content safety, rate limits, fail-mode policy table
  • Sovereign sanitizer designed for non-collection rather than redaction
  • Plan-Preview-Approve-Execute flow + Git-vs-Azure drift detection + dry-run mode
  • Anchored in a 27-section enterprise-sovereign technical specification
Prediction Markets & Derivatives R&D

Algorithmic Trading System

Multi-strategy execution stack for Kalshi prediction markets and IBKR ForecastEx

OUTCOME FOR THE TRADER Trade autonomously — losing strategies dropped before they cost real money. Every strategy in the stack: MULTI-STRATEGY Many side-by-side, independent Each its own kill switch. Independently sized. Independently validated. RISK-GATED Validated before going live Five-gate harness. Deflated Sharpe. Asymmetry-aware sizing. P&L ATTRIBUTION CFO-grade per-strategy P&L Risk-officer auditable. Per-series attribution. Live Streamlit board. ALWAYS-ON Runs while you sleep NTP-verified clock. 2FA auto-restart. 24/7 execution.
What It Does
  • Trades prediction markets and binary derivatives autonomously
  • Runs many strategies side-by-side — each individually validated, each with its own kill switch
  • Drops losing strategies before they cost real money; keeps the winners on automatic
  • Produces full per-strategy P&L attribution that a CFO or risk officer can audit
Under the Hood
  • Order management against Kalshi and IBKR ForecastEx with NTP-verified clock and 2FA auto-restart
  • Strategy library: maker market-making on S&P / Nasdaq / gold / silver / copper, 97%-WR crypto pin, weekly-range accumulation, cross-coin alignment with VIX/DVOL gate, LLM-tournament sports book
  • Five-gate validation harness per series: breakeven WR, asymmetry-aware sample size, tier sizing, deflated Sharpe, paired-bootstrap CI
  • Orderbook collectors, tick recorders, and a Barchart commodity scraper across 14 tickers and 7 page types
  • Streamlit live P&L dashboard with full per-strategy and per-series attribution
  • Circuit breakers: consecutive-loss kill, weather-extreme veto, 5-per-hour rate cap
Construction Tech · Multimodal AI Estimator

Construction Estimator

Multimodal AI Estimator

OUTCOME FOR THE BUILDER A 30-minute walkthrough becomes a tiered Investment Range overnight. Every estimate ships with: MULTIMODAL INTAKE Video → structured scope Claude 4.7 Vision. Deepgram Nova narration. Rooms, fixtures, finish-grade. 24 EXPERT AGENTS One per trade + region + code Plumbing, electrical, framing. Validator + Devil’s Advocate. Consensus threshold gating. 3-TIER PRICING Whirlpool / KitchenAid / Sub-Zero RSMeans + Home Depot Pro. Ferguson + manufacturer feeds. Fort Collins regional cal. FEEDBACK LOOP WhatsApp voice corrections Per-tenant memory layer. Sharper every project. ~$8–14 per estimate.
What It Does
  • Turns a 15–30 minute walkthrough video into a structured scope — rooms, fixtures, rough dimensions, condition assessment, finish-grade — without manual data entry
  • Issues three concurrent estimates (Whirlpool / KitchenAid / Sub-Zero tiers) so clients pre-segment by budget before the site visit, filtering lookie-loos and accelerating bid turnaround from weeks to overnight
  • Calibrates pricing to Fort Collins / Loveland / Greeley with RSMeans regional multipliers, live Home Depot Pro and Ferguson pulls, and manufacturer feeds for premium-spec brands
  • Captures WhatsApp voice corrections from the builder and turns them into a per-tenant pricing memory that sharpens with every project run
  • Anchors the UX to the premium buyer (Sub-Zero / Wolf / Thermador) with lower tiers visually de-emphasized — protects the builder’s brand position
Under the Hood
  • LangGraph durable state machine on top of the Claude Agent SDK, with MCP as the portable tool layer so contractors 2–5 plug in without rewrites
  • 24 domain-expert agents (one per trade plus regional, code, and coordination specialists), grounded in model code books, manufacturer specs, supplier catalogs, RSMeans, and the builder’s own bid corpus over time
  • Multimodal scope analyzer: Claude 4.7 Vision (primary) with Gemini 2.5 Pro fallback for long videos; ffmpeg frame sampling with scene-change detection; Deepgram Nova for narration transcription
  • Pricing layer combining RSMeans Online baselines, Home Depot Pro and Ferguson live retail, premium-tier manufacturer feeds, and a regional cost multiplier calibrated to Northern Colorado
  • Validator Agent and Devil’s Advocate Agent argue against the first pass until confidence crosses threshold; Langfuse for end-to-end observability
  • ~$8–14 per estimate at current model prices, driving toward $4–7 with aggressive prompt caching and a Gemini fallback for long-tail frame analysis
Federal Systems · Defense

Military Equipment Contractor

Configuration-driven CRM for a military equipment contractor

OUTCOME FOR THE CONTRACTOR One auditable pipeline. No more spreadsheets and email threads. Every deal in the pipeline: AUDIT TRAIL ITAR-ready from day one Role-gated access. Review queue built in. Persistent audit log. SMART ROUTING Right team, every time 4 ICPs covered. Direct + program offices. FMS + teaming routed. ENG PRIORITY Scarce capacity to top-value work 72-priority scoring. No squeaky-wheel chaos. Engineering queue tracked. ROLE CLARITY Sales, eng, ops each see their job Role-based dashboards. Live: 165 accounts. 40 active opportunities.
What It Does
  • Replaces a stack of spreadsheets and email threads with one auditable pipeline
  • Routes deals to the right team based on customer type and compliance gate
  • Ranks engineering priorities so scarce technical capacity goes to the highest-value pursuits
  • Gives operations the audit trail it needs for ITAR review
Under the Hood
  • Configuration-driven schema — 13 entities, every field, rule, and stage declared rather than hardcoded
  • Four ICPs covering direct end users, program offices, FMS, and teaming arrangements
  • Engineering Queue with custom 72-priority scoring matrix
  • ITAR Review Queue and Info Log → Lead conversion workflow
  • Role-based dashboards split for Sales, Engineering, and Operations
  • Live scale: 165 accounts · 81 contacts · 40 active opportunities · 38 logged activities
Industrial & Energy Services Portfolio · 5-State Region

Private Equity Firm

Agentic business intelligence platform for an industrial portfolio company group

OUTCOME FOR THE BD TEAM Find projects months before public bid windows. Every opportunity surfaced: EARLY SIGNAL Months before public bid windows 14 automated scanners. EIA, SAM.gov, EDGAR. BLM solar GIS, more. PORTCO ROUTED Auto-routed to the right portco 4 portfolio ICPs. NAICS + geography fit. Size + technographics. LIVE SCALE 1,200+ companies 4,800+ contacts 42 active projects. 35 verified subs. Welding, waterjet, laser. CROSS-PORTFOLIO One map, not ten spreadsheets Five-state coverage. Project → co. → sub. Cascading filters.
What It Does
  • Finds energy and construction projects months before public bid windows
  • Identifies the EPCs, developers, and subcontractors building each project
  • Routes every opportunity to the right portfolio company automatically
  • Lets the BD team work from one map instead of a stack of spreadsheets and news alerts
Under the Hood
  • 14 automated scanners: EIA, SAM.gov, USAspending, SEC EDGAR, BLM solar GIS, state procurement portals, industry news
  • Scoring engine on NAICS match, geography, dollar value, and project type
  • Four portfolio-company ICPs scored against industry fit, size, geography, technographics
  • Visual mapping interface: Projects → Companies → Contacts → Subcontractors with cascading filters
  • Live scale: 42 active projects · 1,200+ companies · 4,800+ contacts · 35 verified subcontractors
  • Five-state coverage across welding, waterjet, laser cutting, and filtration portfolio companies

Have a project in mind?

I take on a limited number of engagements at a time. Let's talk about whether yours is a good fit.

Start a Conversation