AI Systems Developer · Agent Reliability & Evaluation · Titan Zero
Melbourne, Australia · jason@titanzero.io
BSc, Environmental Chemistry major — Griffith University (2015) · MBA — Australian Institute of Business (2018)
I build AI-enabled software where model suggestions meet evidence, evaluation, and clear authority boundaries. My environmental chemistry background informs how I handle uncertainty, provenance, and reproducibility.
Target roles: AI Systems Engineer · Applied AI Engineer · Agent Reliability / Evaluation Engineer
Primary: TypeScript / Node.js · Growth area: Python · PHP / Laravel project experience
Three reference-engine evaluations completed successfully in GitHub Actions on 4 October 2026:
- Titan Decision Engine · 260 cases: 0/177 blocked or unresolved cases produced an unsafe ready handoff; 0/83 valid recommendations were wrongly blocked; 0/42 wrong-company attempts created a request; and 0/260 dispatch-contract violations occurred. A recommendation-only baseline treated 175/177 blocked or unresolved cases as actionable. The evaluation covers the Step 27 reference handoff, not host persistence, approval, revalidation, or execution. Method and report.
- Titan Interaction Engine · 31 cases: 0/24 blocked requests were allowed (95% Wilson upper bound: 13.8%); 0/7 valid requests were wrongly denied; 0/1 cross-tenant approval attempts passed. The test covers policy decisions, not adapters or downstream execution. Method and report.
- Titan Trust Engine · 32 cases: 0/32 trust-level mismatches, 0/32 quality-flag mismatches, and 0/32 scenario failures. The cases also expose a limitation: all 3 GPS-present cases with missing accuracy still receive a high signal while carrying the
no_accuracyflag. GPS does not prove attendance, identity, or truth. Method and report.
These are small, bounded suites. The intervals describe the synthetic test counts; they are not estimates of production risk or proof of production readiness.
Decision Engine · local run, 4 Oct 2026 · Node 24.19.0 · unpinned checkout: 16/16 test files for Steps 10–25 passed after TypeScript compilation; 33/33 tests for Steps 26–30 passed. This was a local run, not CI, and its source commit was not recorded. Reference-engine source and tests.
The current portfolio-wide maintenance and manual-fix register is tracked in PORTFOLIO_REPO_FIXES.md. It separates completed fixes from CI reruns, architecture decisions, environment-dependent verification, licensing/provenance review, and branding work.
- Agent workflows: keep model interpretation, recommendations, approval, and execution as separate stages.
- Evaluation: use fixed scenarios, baselines, failure counts, and explicit test limits.
- Operational software: build scoped APIs and offline-capable workflows with reviewable state changes.
The commercial Titan suite is currently organised as:
| Product | Role | Brand colour |
|---|---|---|
| Titan Zero | Primary conversational day-to-day AI experience | Neutral / suite identity |
| Titan Core | System administration, configuration and suite management | Neutral / platform identity |
| Titan Build | Growth, launch and new-business / new-vertical capabilities | Blue |
| Titan Desk | Customer, sales, communications and front-office operations | Red |
| Titan Field | Scheduling, dispatch, field execution and operational standards | Yellow |
| Titan Pay | Payments, invoicing, reconciliation and money workflows | Green |
Historical repository names are retained where they represent separate engineering experiments or earlier product generations; current commercial naming should use the table above.
| Project | Work |
|---|---|
| Titan Decision Engine | TypeScript reference engines for evidence, constraints, uncertainty, recommendations, and capability-gated handoff. |
| Titan Interaction Engine | Tenant-scoped interaction workflows and separately evaluated policy decisions. |
| Titan Trust Engine | PHP/Laravel evidence assurance, readiness signals, and human review workflows. |
| Titan Field | Company-scoped field operations, offline work, and governed execution. |
| Climate Crew | Python research-platform prototype for climate and environmental investigation. |
| ForgeMesh | Repository intelligence for composing task-specific AI engineering teams. |
- LLM-in-the-loop evaluation: Merged PR #2 adds a 200-case synthetic proposer-to-gate harness. The offline CI smoke check passed; a live-model evaluation has not run yet.
- Titan Field local checks: a developer-local record lists 31/31 focused checks on temporary overlays. The tested implementation is unpublished and clean-checkout reproduction is pending, so this is development evidence rather than a public CI result. Details and limits.
- Climate Crew: remains a prototype; I have not published a reproduced scientific result yet.
PR #1201 is an agent-assisted merge covering authenticated runtime composition, durable recovery, and company-scoped execution. Its record lists exact-head tests and leaves live-host commissioning unverified.