pl8ypus

Tanya / AI-assisted self-assessment / 5 November 2026

From AI-assisted builds
to connected delivery.

Greg’s current evidence combines delivered applications with a working agent harness: scoped local coding, durable workflows, prepared context, reviewed memory, tracing, and separate source review through to application integration.

Emerging Forward Deployed Engineer Connected lab delivery Confidential external use

Greg uses Tanya, an AI-assisted workflow, to review his own evidence. This page reflects the build record and Greg’s delivery context through 5 November 2026. It is a self-assessment, not an independent endorsement or audit.

Current assessment

Stronger applied engineering evidence. A longer operating record is still needed.

The assessment remains Emerging Forward Deployed Engineer. Greg can show both useful software delivery and the engineering needed to coordinate AI-assisted work: contracts, bounded execution, tests, review, integration, and an inspectable record.

Demonstrated

A connected delivery process

The selected lab stack has carried prepared context through local implementation, native checks, separate source review, and application integration. A reviewed contribution reached an application without lead edits to the worker’s accepted implementation.

Confirmed by Greg

Confidential external use

A major organisation is using the lab on a significant project. That adds an adoption signal. The organisation, sector, project details, and outcomes remain confidential; this does not establish project-wide production status.

The distinction that matters now is between demonstrated supervised delivery and sustained operating responsibility. Measurable customer outcomes, repeated use, support, and recovery still need a longer record.

AI lab / Current engineering evidence

The tools now form a working delivery harness.

The updated lab story describes the current architecture and actual worker tools. For this assessment, the value is what those components allow Greg to control, inspect, and hand over.

Implementation

LangChain + Deep Agents inside Caraval

The selected model/tool graph works through Caraval’s declared file access, source hashes, fixed checks, finite budgets, and stop conditions. Local Coder Next workers implement bounded tasks; the hosted Codex lead coordinates preparation, review, and integration.

Workflow continuity

LangGraph + SQLite

Persisted stages connect intake, dispatch, review, and handover. Completed workflow state has been reopened without new model calls. This is concrete continuity evidence; wider interrupted-work recovery remains a separate requirement.

Task context

Docling + MCP / FastMCP

Supported local HTML is converted to recorded Markdown, read through allowlisted tools, and frozen into task inputs. The relevant engineering is traceable context preparation and provenance, with the worker’s file permissions preserved.

Memory discipline

Mem0 + Qdrant

Eligible lessons are checked against their original authority before retrieval is used. Recent tasks correctly returned no eligible lesson. The retrieval path is implemented; causal improvement in quality or effort still needs comparative evidence.

Operational visibility

Dashboard, Studio, Phoenix + LangSmith

Assignments, dependencies, role signals, handovers, host resources, and instrumented traces are visible. OpenInference supports local tracing; selected metadata reaches LangSmith. Observation distinguishes recorded history from active work and does not dispatch tasks.

Acceptance judgement

Native checks, promptfoo + source review

promptfoo exercises positive and negative controls on completed evidence. Native checks test the candidate. Separate source review remains essential: a fresh November CLI candidate passed tests but was rejected for UTF-8 defects; the lead’s correction was separately requalified and reviewed.

These components have recorded use in selected supervised tasks. The strongest evidence is the complete handover path and the willingness to reject passing code when review finds a contract defect. See the current harness diagram or inspect the worker tool surface.

Supporting portfolio

The wider delivery record still matters.

The lab extends an existing application and enterprise workflow record. These are supporting examples from the published build stories; their scope and delivery states remain distinct.

Enterprise campaign delivery

Campaign Copilot

A built campaign orchestration system deployed with a global company, with integration APIs around governed intake and build handoff. This supports the customer-workflow and delivery case. Read the build story.

Enterprise application delivery

Hourloom

A completed enterprise deployment for time and project workflows, with durable timers and report reconciliation. It adds application state and persistence evidence. Read the build story.

Accepted and in use

Marketing Form Protection

The enterprise form-protection workflow has been accepted, deployed, and put into use. It connects an operational marketing problem with controlled forwarding and human review. Client identity remains private. Read the build story.

Cloudflare application delivery

Client Portal

The published case study documents a live application with client and administrative workflows, stored records, attachments, and activity evidence. Its deployment and security-review scope belong to that case study. Read the case study.

Desktop engineering

TargetSize

The Windows desktop utility adds .NET/WPF, FFmpeg integration, deterministic planning, dependency handling, and a documented debugging trail. It broadens the engineering record beyond web applications. Read the build story.

Workflow and governance design

Tanya Build Cockpit + Translation AI

The cockpit contributes structured task state, retained context, and review gates. Translation AI contributes governed output and human-review workflow design. Each build story defines its own implementation and rollout scope. View the cockpit · View Translation AI.

Fit for customer-facing engineering

The strongest signal is judgement across the whole task.

Enterprise marketing experience helps Greg connect the technical work to the operator’s problem. The connected lab makes that judgement visible in the way tasks are scoped, checked, and adopted.

Problem definition

Turn workflow knowledge into a contract

Campaign operations, form quality, translation review, and client support provide concrete requirements. The current task packets make ownership, accepted dependencies, required inputs, and checks explicit.

Delivery control

Coordinate tools without surrendering acceptance

Worker scope, repair limits, source identity, and separate review carry through the framework integrations. A model’s finish and a native PASS have defined meanings; application integration remains its own decision.

Operator ownership

Make progress and failure inspectable

The dashboard and traces expose the work behind a handover. Required-input coverage and retained rejected candidates help the lead identify what still needs attention before the next assignment.

Credentials and communication

Supporting signals alongside the delivery record.

The existing credentials and speaking record support the technical and customer-facing parts of the profile. The linked pages retain the detailed records.

Formal learning

Recorded AI credentials

The certification page records Greg’s OpenAI practitioner credentials, Codex Expert badge, and other AI learning. These complement the implementation evidence. View the credential record.

Field communication

Private partner speaking selection

The speaking page records Greg’s selection to explain practical AI delivery in a private professional setting. It supports communication experience, with event details kept private. View the speaking record.

Remaining evidence

The next proof is sustained operation.

The lab now has workflow state, tracing, a dashboard, and controlled delivery. The remaining questions concern breadth, effectiveness, and operational responsibility over time.

Repeatability

More accepted work across task families

Repeat useful application contributions with clear owners, accepted dependencies, source review, and full integration checks. A small set of accepted tasks is not yet a general worker-reliability measure.

Impact

Measured outcomes and effort

Record delivery time, review and correction effort, quality, cost, adoption, and customer outcomes where disclosure is permitted. Test totals and local timings alone do not establish return on investment.

Recovery and support

A longer operating record

Extend completed-workflow reopen evidence to broader interruption scenarios and verified off-device recovery. Show support ownership, incident handling, controlled change, and maintenance during continued use.

Memory and automation

Demonstrated benefit before broader autonomy

Compare retrieval with suitable controls to measure memory benefit. Validate broader unattended operation separately. Neither a live role map nor installed frameworks establishes an autonomous delivery team.

The current positioning is an Emerging Forward Deployed Engineer with enterprise workflow knowledge, application delivery experience, and an implemented agent harness. The next assessment should be driven by repeated accepted delivery, measurable outcomes, and sustained operational ownership.

Discuss an applied AI delivery problem.

Enterprise workflows, connected agent tools, and software delivery with clear review and ownership.

Explore the AI lab