Products I've designed, built and run in production.

Skopos

AI-native vendor risk intelligence for cybersecurity teams.

A multi-tenant vendor management platform for TPRM consultants and their clients. It covers the whole vendor risk lifecycle, from intake through due diligence, evidence, scoring, findings, remediation and an audit-ready report. One SOC 2 report can be shared across many clients while every assessment stays isolated to its tenant.

• Vendor registry shared across tenants, assessments isolated per client
• Review workflows from intake to sign-off
• Questionnaire engine with framework-mapped templates
• Evidence management with AI document and SOC 2 analysis, plus a locally fine-tuned classifier for confidential reports
• Risk scoring against each tenant's risk appetite
• Audit packages exported in one step
• MFA and RBAC: TOTP enrollment, five deny-by-default roles
• Tenant provisioning with default roles and risk models
• Agent workflows: tenant-defined AI workflows with human approval gates
• Agent activity dashboard with a full audit trail of every AI action

Agentic Phone

A 24/7 answering service that sounds like the front desk.

Each business gets its own phone number, voice, greeting and call script. The agent holds a real conversation, recognizes emergencies, reads back callback numbers, and emails the owner a transcript with a priority flag after every call. Live demo lines cover plumbing, electrical, medical, legal intake, real estate and a car dealership.

• Per-tenant agents: voice, persona, tools and prompt per business
• Emergency triage, e.g. a gas smell versus a slow drain
• Post-call pipeline: transcript, call log and analytics
• English and Spanish callers
• Contact Flow routing by dialed number
• Live in under an hour for a new business

CleanOps Cloud

Run a whole cleaning business from one calm place.

Software for residential cleaning businesses and commercial cleaning teams. It covers scheduling, quotes, payments, payroll-ready time tracking and a customer portal, with separate experiences for owners, cleaners and customers. Plans scale from a solo cleaner to multi-crew operations, with no add-on modules.

• Scheduling that fills the calendar
• Quote, bill and collect in one flow
• Crew management with time tracking for payroll
• Customer portal for bookings and payments
• Native iOS app for cleaners in the field
• Four roles: super admin, owner, cleaner, customer

WalkaroundAI

Turn listing photos into a cinematic, narrated video.

Paste a listing URL or upload photos and get a narrated walkaround in minutes: an exterior orbit then the interior for vehicles, or room-to-room for homes. Built for car dealers, real estate agents, and boat and RV sellers. Pricing is per video, starting at $9.99.

• Listing URL in, video out, no filming
• Cars, homes, boats and RVs
• AI narration written from the listing details
• Watermarked preview before purchase
• Dealer and agent packs for volume
• Rendered on owned hardware to keep unit cost low

From issue to production without a hand-off meeting

Seventeen specialist agents plan, build, verify and operate the software, under a CTO agent that assigns the work and merges the results.

Each GitHub issue moves through a fixed pipeline: Product Owner, UX Designer, Developer, Security Reviewer, Tester, then staging deploy and QA. GitHub labels act as the state machine, and each agent writes a structured hand-off that the next one reads. Any agent can send work back for rework. Agents never merge their own pull requests: only the top-level orchestrator merges, one PR at a time, after checks pass. Separate long-running agents handle defect intake, production support, SRE and sprint execution around the clock.

• 3,337pull requests merged
• 3,700issues closed
• 4,838commits across 14 repos
• ~12 / dayPRs merged on Skopos since March 2026

• 17 agent roles on Skopos, reused across Skopos Intelligence and Agentic Phone
• 26 orchestration skills: sprints, CI repair, staging deploys, QA, prod monitoring, error triage, overnight runs
• Up to ~20 parallel tracks, one git worktree each, with risky changes isolated
• 4 always-on agents: defect intake, prod support, SRE and a 30-minute sprint loop
• Shared memory across 8 projects so agents keep context between sessions
• Security review gate on every change: OWASP, tenant isolation, IAM least privilege
• 10-worker Playwright regression suite with an AI vision pass

A 180B model on one consumer GPU

Qwen3.8-Flash-Next, a 180-billion-parameter mixture-of-experts model, served locally through Strata.

I deployed four builds of the model (IQ3_S, Unsloth 4-bit, Coder and Swift 1.5) and wrote a benchmark harness covering code generation, reasoning, math, plain chat and tool calls, with cold and warm runs. The report compares Strata against LM Studio (llama.cpp) on the same model and on a 27B dense model, and traces speed back to its cause.

• 188 tok/speak generation, Coder build
• 16.7×faster than LM Studio, same model
• 3.0×faster than LM Studio on a 27B dense model

• Expert-cache hit rate identified as the main speed lever (r = 0.75 across 20 runs)
• Weak spot found: short tool calls drop cache hits to 48–68% on full-expert builds
• No warm-up penalty: cold and warm runs within 1–5%
• OpenAI- and Anthropic-compatible API on localhost for agents and apps

Headless ComfyUI render workers

Image-to-video generation with Wan 2.2 14B, packaged as a container that runs locally or on AWS Batch.

The worker pulls jobs from a local folder or S3, renders each clip through ComfyUI's API, caches good clips so only rejected shots re-render, and finishes the video with frame interpolation, narration and a watermarked preview. Several workers share one queue through lease files with heartbeats and takeover.

• ~154 sper 5-second 720p clip, models cached
• 28 of 32 GBpeak VRAM while sampling

• Seed-reproducible clips, verified frame by frame
• Resource metrics recorded on every run
• Blackwell GPU support worked out (CUDA 13 / PyTorch cu130)
• Same image locally and in the cloud

A private model for SOC 2 review

A fine-tuned 7B model that sorts vendor security documents before analysis, trained and served entirely on local hardware.

Vendors send everything from SOC 2 Type 2 reports to marketing whitepapers, and an analyzer needs to know which it has before it can judge controls. I built the full pipeline: inventory a corpus of 376 security PDFs, extract the text, hand-label documents into a 16-type taxonomy (SOC 2 Type 1 and 2, SOC 3, bridge letters, ISO, HITRUST, PCI, pen tests, DPAs, BC/DR plans and more), build supervised fine-tuning sets, train LoRA adapters on the RTX 5090, evaluate against a held-out test set with a confusion matrix, and quantize the result for Ollama. Client reports never leave the machine.

• 376security PDFs in the corpus
• 333labeled examples, split train / val / test
• 16document types in the taxonomy

• Held-out evaluation with per-class confusion, never trained on test data
• Structured JSON output: type, company, confidence, reason
• 4.7 GB quantized model served locally through Ollama
• Next: balance the classes, then move from classifying to control-level review

Infragil

Secure AI adoption for the businesses that can't afford to get it wrong.

The company site. It sells advisory work to healthcare and regulated organizations adopting AI, and it is the front door for Skopos.

• AI readiness and governance assessments
• Executive AI security and risk advisory
• Virtual CISO and TPRM program support
• AI automation and agentic AI delivery

The CISO Board

A collective of security leaders offering vCISO and vCXO services.

A site for a group of senior security executives who provide fractional CISO leadership. Services are tiered by an organization's maturity and risk profile, from baseline guidance for growing businesses to embedded executive leadership.

• Tiered vCISO plans by maturity and risk
• Board meeting representation
• Continuous compliance monitoring
• Confidential consultation intake form