Product · Partner ecosystems · AI systems
Enterprise experience.
Practical AI systems.
I bring 10+ years in product management, partner ecosystems and commercial operations to building tools for the work I know.
I define the architecture, direct AI-assisted implementation, and personally test and validate the result. The work below shows the systems, the decisions and the evidence behind them.
BatonHand-off CLI$ baton pass builder "the ads backfill finished — 412k rows, want the diff?"01K2F7Q8VZ3NBRJ4WYD6XM0AHC
# the id comes back from the hub, not the client.
# a send that could not be recorded fails, loudly.Placeholder · CLI from the Baton README
BespokeNative chat capture
Selected builds / 2026Move across to browse
Index / Case studies
Selected builds
Bespoke · DIA
Case Studies & Builds
01 / From problem to systemDIA
A questionnaire becomes a scored opportunity and three useful documents.
A deterministic pipeline handles parsing, validation, scoring and document generation. A Gemini agent uses those functions as tools.
I brought together the workflow, interface, document generation, testing and deployment for the Google for Startups AI Agents Challenge.
Workflow illustration · not a product screenshot
Design decisions & what to inspect
The deterministic path makes the scoring and generated documents inspectable without relying on a model response. The agent can call the same functions rather than invent a separate scoring process.
In the demo, use Load Sample to inspect the synthetic questionnaire and generated documents. The live agent is a separate path; a working deterministic demo does not establish model availability.
The public repository includes parser and scoring logic, fixtures, tests and an evaluation harness. See the dated build record ↓
Blackboard
A visible record of work across people and agents.
A local viewer for file-based events: handoffs, shared artifacts and activity across the fleet. I use it in my own environment and refine it through practice.
The record remains readable even when the viewer is down.
Events are plain JSON files. This describes the local implementation; there is no public demo here.The build record
02 / Selected evidenceA few dated implementation steps. Open a note for the scope and source behind it.
06–09 Jun
2026DIA: from public demo to capture toolingPublic commits and a scoped historical test record
- 06 Jun · 19:43
- Initial public demo. The commit contains the parser, scoring, document generation, web interface and tests. Commit
- 09 Jun · 01:01
- Capture integration. The original integration note records 26 passing tests across three named test modules. This is a historical result, not a current full-suite total. Original note
- 09 Jun · 02:37
- Capture made opt-in. The implementation gates OBS capture behind an explicit setting and catches failures so they do not break the intake response. Implementation
Times shown in EDT (UTC−04:00), from the original commit records. No elapsed-time or submission claim is inferred.
27–28 Jul
2026Amazon Ads: a queryable reporting pipelineAPI retrieval → durable storage → console-export ingestion
- 27 Jul · 23:53
- Report retrieval. A v1 API client creates or resumes a report, waits for completion, then downloads it.
- 28 Jul · 00:41
- Durable data layer. Parquet snapshots feed a rebuildable DuckDB store and a read-only query interface.
- 28 Jul · 00:57
- Console-export support. An additional ingestion path normalizes exported report columns and handles year-month periods.
EDT (UTC−04:00). Implementation checked against the corresponding local repository revisions. These are engineering milestones; account data, client outputs and commercial performance are excluded.
08 Oct
2026Bespoke: repeatable desktop actionsLocal implementation · verified step recipes and durable targets
Work on remembering a step after its result has been checked, then replaying it against a durable target. Replay follows the operator’s selected approval mode.
Internal implementation record; source remains private. This is not a public release, a client deployment or a measured productivity result.
Tools I work with
03 / The supporting systemsBespoke
Operator-controlled reading, voice interaction and desktop actions. A local tool for moving between the written word, conversation and repeatable steps.
Local implementation- speak
- Read text with playback controls.
- ask
- Answer a prompt, then speak the response.
- live
- Stream microphone audio and native audio replies, with interruption handling.
- cursor
- Desktop actions with durable targets and verified step recipes.
Scope & implementation
The repository implements both turn-based ask and a separate real-time live mode. The latter streams audio over a websocket and flushes playback when interrupted.
Desktop replay follows the operator’s selected approval mode. Verified step recipes and durable targets describe the local implementation; they do not establish independent operation or client deployment.
This portfolio describes implementation work, not a new end-to-end microphone or latency test. It is not presented as a full accessibility screen reader.
Baton
Routing and coordination between agents, supporting how work moves through the wider environment.
Working locallyAgent Eggs
Portable context and capabilities, packaged so useful knowledge can travel between environments.
Reusable context packagesResearch & writing
04 / The thinking behind the workdPKL is the foundation of how I think about AI systems. These four connected works explore knowledge, execution, the operator’s role and human autonomy.
The Deterministic Personal Knowledge Layer (dPKL): Agents as the Interface to Deep Execution
A local-first approach to making working knowledge portable and queryable, with agents acting as an interface to execution.
What Expectations? AI, Telecommunications, and the Tool It Always Was
AI as infrastructure and a tool, with human judgment shaping how its capabilities are used.
Agents Are Your New Mouse and Keyboard: An Operator’s Doctrine for Driving Agents
An operating practice for directing agents through command-line tools, documentation, reusable capabilities and human validation.
What Autonomy Is, and Why It Remains Exclusive to Humans
A philosophical argument for keeping purpose and responsibility with humans, and designing agents for bounded, reliable execution.
The ideas and operational experience are mine; each work discloses agent-assisted prose and formatting. Full texts are not published here.
Where the judgment comes from
05 / Enterprise experienceProduct lines, partner programs and commercial operations taught me to account for the messy parts: incomplete information, handoffs and competing priorities.
That experience shapes the problems I choose, the systems I design and what I check before calling something useful.
Put a workflow to work.
06 / Work with meHave a recurring workflow that needs attention? I work on business automation through BASH. Start with one task, agree to the scope and acceptance checks, and then build from there.
navy@bashsystems.net