DBEvidence-bound systems
Menu

Public résumé / evidence edition

William “Drew” Baker

Software engineering · Geospatial evidence · Climate-relevant systems

Software engineer building inspectable systems, deterministic software authorization boundaries, and geospatial evidence workflows for high-consequence settings. Public projects show bounded model evaluation and release testing; historical coursework adds energy-policy context.

Contact. For professional inquiries, use LinkedIn. Direct email, phone, and location are intentionally omitted from this public résumé.

01

Hiring order / implemented and evaluated work

Selected project evidence

Flagship 01BurnLens reached a public, complete Phase Six portfolio release: v0.56.0-baseline-first-portfolio-release at commit e2e0b778; the later evidence snapshot is a741111d, four commits after the release.

BurnLens

My role
Drew set the portfolio thesis, target audience, use boundaries, owner stop conditions, and publication direction, and owned the human decisions; Codex was assigned technical, product, and reliability direction within that owner-defined envelope.
What I built
The public release had to remain useful while excluding credentials, private owner responses, private logs, and machine-local paths, without implying official or operational capability.
How I tested it
Reliability includes recognizing when evidence is insufficient and making that stop reproducible.
Result
BurnLens reached a public, complete Phase Six portfolio release: v0.56.0-baseline-first-portfolio-release at commit e2e0b778; the later evidence snapshot is a741111d, four commits after the release.
Current boundary
BurnLens is experimental portfolio evidence, not official wildfire information, emergency guidance, or operational decision support.
Read the case study Inspect pinned evidence (opens in a new tab)
Flagship 02Verified synthetic testbed with a public v0.0.20 release and zero real systems connected.

Runbook Sentinel

My role
Repository author and release owner for the pinned public checkpoint.
What I built
A dependency-free synthetic control system that separates reasoning, approval, policy, execution, state, and audit.
Methods and technologies

Pinned standard-library runtime and release primitives.

  • Python 3.12+Runtime, evaluation harness, verifiers, API and bounded diagnostic-tool server, and packaging
  • SQLiteSynthetic state, proposals, approvals, idempotency, execution, and audit persistence
  • JSON records and chained event logsContracts, evaluation artifacts, traces, and evidence bindings
  • JSON-RPC over stdioDiagnostic and read-only tool surface with no approval or execution tool
  • Python zipappDependency-free release artifact with exact allowlist and hashes
  • Loopback HTTP and server-rendered HTMLLocal operator, API, and dashboard surfaces
How I tested it
Exact synthetic trajectories and final synthetic system states, split-aware coverage, fail-closed contracts, package parity, and real-surface verification.
Result
The pinned v0.0.20 synthetic release passed its exact deterministic gate and excluded the weaker local-model candidate.
Current boundary
A released synthetic testbed with explicit infrastructure, identity, trace, security, and evaluation limits.
Read the case study Inspect pinned evidence (opens in a new tab)
Supporting implementationReviewed prototype represented by a frozen public reviewer snapshot.

Quest Craft

My role
Product and system designer and release owner directing AI-assisted implementation, evaluation, correction, and approval.
What I built
One constrained generation request produces three structured paths while local checks and the adult authority rail remain outside the model step.
How I tested it
A fixed release suite of 12 synthetic scenarios, three retained rows each, human-scored cells, local privacy rejections, and corrective reruns.
Result
The reviewer snapshot retains 36 rows: 33 model generations, three local privacy rejections, six failed or superseded attempts, and two behavior corrections.
Current boundary
A bounded public reviewer snapshot, not the private canonical implementation or a general safety result.
Read the field note Inspect pinned evidence (opens in a new tab)
02

Professional record

Experience

Apr 2024 – Apr 2026

Gen AI Associate

Innodata

  • Annotated and evaluated 200–500 AI prompt-and-response tasks weekly across multiple projects for model-training and evaluation workflows.
  • Worked with a distributed team of more than 35 people to identify failure modes, refine prompts, and align outputs with project requirements.

Nov 2021 – Mar 2024

Technical Content Writer

Independent contractor

  • Wrote and edited more than 300 technology articles, including energy and residential-solar explainers.
  • Translated technical and energy topics into accessible material for broad, non-specialist audiences.