Turns operating problems into systems he then runs, directing parallel agents and never accepting output without an independent check.
Structured profile submitted by its owner. Overflow validates and normalizes the payload, but does not independently verify claims derived from local source history.
01 / Work profile
Operator Engineer evidence map.
Labels organize the proof across technical chops, business know-how, and good judgment; the cited work is the claim.
Repeatedly got substantive change into live use — a signed and notarized desktop build with production payments activated, a platform ownership transfer to an external stakeholder, and a single sign-on rollout reaching a real external user.
Why ★★★5 qualifying arcs · 4 systems · 800 days
Technical chops★★★T-01
Prototyper
Repeatedly turned ambiguous ideas into artifacts a real user could try, from a document-extraction prototype through an assessment platform, a subscription application and a browser game used as a lead magnet.
Why ★★★4 qualifying arcs · 4 systems · 340 days
Technical chops★★T-06
AgentOrchestrator
Decomposed work across named workers with separate goal documents, kept lead and doer sessions apart, gated them with pass/fail checkpoints, and reassigned a stalled worker with an explicit file-ownership handoff.
Next starThe same orchestration pattern demonstrated on a system also worked in at least 90 days earlier; qualifying evidence currently begins 30 Jun 2026 because older transcripts are no longer readable.
Technical chops★★T-05
ContextEngineer
Keeps durable context outside the conversation through reusable skill files rewritten after each production cycle, per-workstream goal documents read before work starts, and a nightly job that folds lessons back into his own instruction files.
Next starContext handoff artifacts demonstrably reused on a second system across a gap of at least 90 days, rather than within a single dense month.
Technical chops★★T-02
FrontendCrafter
Commissioned a formal design system of tokens, radii and type and had parallel agents adopt it across two codebases, and reviews interfaces at named breakpoints using element-level comments carrying selector and pixel position.
Next starInterface work separated by at least 90 days; every qualifying arc currently falls between 24 Jun and 24 Jul 2026. Accessibility judgment is not yet evidenced at all.
Technical chops★★T-04
SystemsArchitect
Made structural calls that survived into implementation — schema-versioned profiles required to keep older records working untouched, and a scope boundary placing an extracted SDK in a new repository rather than modifying its source system.
Next starA consequential architecture decision in a system also operated at least 90 days earlier; readable transcripts for these arcs begin 11 Jun 2026.
Business know-how4 earned
Business know-how★★B-01
WorkflowArchitect
Mapped a real client-review cadence into a scheduled brief merging mail, chat, meeting transcripts and ticket data, defined what fresh, partial and missing data mean, and restricted capture to meetings the company owns.
Next starA second workflow of comparable depth shown still running at least 90 days after it was designed.
Business know-how★★B-02
ValueTranslator
Owns the commercial layer directly — tiered fees, intellectual-property and renewal terms, and a payback calculator built so it cannot overstate, with the rule that the hard-dollar case must run against the whole system rather than its strongest candidate.
Next starAn economic or pricing call whose realised outcome can be pointed to at least 90 days after it was made.
Business know-how★★B-04
AdoptionOperator
Designs the handoff rather than stopping at delivery — platform and repository ownership moved to an external stakeholder with post-migration functional testing, alongside a written response commitment and meeting cadence governing how engagements run.
Next starEvidence that a handed-off system was still in active use by someone else at least 90 days after the handoff.
Business know-how★★B-03
ClearCommunicator
On judgment-heavy work the brief carries outcome, constraints and a definition of done, and the constraints hold — budget ceilings, no-retention rules and claim-accuracy limits are stated up front and are not relaxed later.
Next starClosing the counterweight: on creative iteration a recurring strand of terse, non-diagnostic feedback leaves the agent to infer direction, and several technical debugging threads end without a stated resolution.
Domain proofDelivery operations ★★★Agent operating models ★★★Services pricing and ROI modelling ★★Pipeline data integrity ★★
Good judgment4 earned
Good judgment★★★J-01
VerificationFirst
Uses an independent check before accepting output — read-only reviewer agents forbidden from editing, a second model brought in to judge the first one's plan, release reviews demanding file-and-line evidence of which one returned a no-go, and renders accepted only against decode checks, frame counts and hashes.
Why ★★★5 qualifying arcs · 4 systems · 336 days
Good judgment★★★J-03
RecoveryOperator
Detected real failures and bounded the retry — a data-loss event recovered by migration, an enrichment source abandoned after repeated outages and replaced only once a single query was proven, and a stalled worker frozen and handed to a fresh session with its files.
Why ★★★4 qualifying arcs · 3 systems · 333 days
Good judgment★★★J-02
TradeoffNavigator
Chooses against named constraints and their alternatives — a spend audit that rewrote the model-routing rules mid-week, an infrastructure choice argued from a no-retention requirement, and a parsing provider swapped only after timeout tuning failed to fix it.
Why ★★★4 qualifying arcs · 3 systems · 333 days
Good judgment★★J-04
SystemsSteward
Keeps live systems tended through recurring defect fixes, a consolidation pass that removed complexity rather than adding it, and scheduled jobs he watches — corroborated by repeat pushes to the same repositories across 27 consecutive active months.
Next starA regression caught by instrumentation rather than by personally noticing something was wrong; reliability problems currently surface reactively, including a scheduled job whose failure was noticed only because an expected email never arrived.
★ Observed★★ Established★★★ DemonstratedStars show proof strength, not skill level.
A consultant-operator who converts business problems into running systems: he sets the metric, the constraint and the acceptance bar, directs agents to build against them, verifies with an independent check, and then keeps the result operating. Ownership spans product decision, commercial terms and post-launch care rather than stopping at implementation.
Share the proof
A LinkedIn-ready evidence card.
Activity, coverage, comparative placement, and concrete work arcs lead. Badge terms appear only as secondary evidence lenses.
Captured agent activity100th percentile · upper third · directional n=1220,272agent sessionsYOU39 lowmiddle 50% · 120–85120,272 high
Captured sessions across the agent-tool windows reported in each assessment.
Claude Code activity100th percentile · upper third · directional n=1219,618Claude Code sessionsYOU14 lowmiddle 50% · 30–35019,618 high
Captured Claude Code sessions in the reported local-history window.
Codex activity78th percentile · upper third · directional n=10374Codex sessionsYOU11 lowmiddle 50% · 60–3392,533 high
Captured Codex sessions in the reported local-history window.
Weekly average across captured span100th percentile · upper third · directional n=12253.9sessions / weekYOU2 lowmiddle 50% · 7–43254 high
A rough intensity signal normalized across the full dated capture window. Tool retention still varies.
Recent agent activity86th percentile · upper third · directional n=8499sessions in 28 daysYOU17 lowmiddle 50% · 43–3821,063 high
Sessions in the 28 days before the assessment.
Active agent weeks79th percentile · upper third · tied with 3 · directional n=812active weeksYOU6 lowmiddle 50% · 7–1212 high
Weeks with at least one captured agent session during the last 12 weeks.
Each curve uses only profiles that report a compatible version of that metric and appears at eight reported values. Baseline activity metrics span all assessment versions; newer recency and GitHub curves appear as their compatible cohorts grow. This compares submitted evidence, not career seniority.
signal received · curiosity is a useful metric
02 / Delivery evidence
Observed work arcs.
Internal delivery and account-operations platform
Ongoing Operation · Established System · Product Operations · System Owner
Merged five separate communication and delivery data sources into a scheduled weekly client-prep brief delivered ahead of review meetings. Defined the coverage vocabulary the system reports against — fresh, partial and missing — rather than accepting an agent-proposed metric. Set a privacy boundary limiting meeting capture to company-owned meetings after initially granting broad access.
Consulting practice web presence and commercial tooling
Live Use · Established System · Experience Interface · System Owner
Rebuilt the homepage by commissioning three parallel design directions and personally selecting among them. Built a payback and financing calculator with optional strategic and enterprise value assumptions, designed so the tool cannot overstate the case. Published case studies only after explicit client approval, and pulled one sector reference that could not be disclosed.
Community proof-of-work assessment platform
Live Use · Early System · Application Software · System Owner
Required schema-versioned profiles to keep older records working unchanged as the format advanced, and that constraint survived into implementation. Directed diagnosis and a recovery migration after an in-flight data-loss event rather than intervening manually. Rolled out single sign-on with account linking and verified-email handling, then commissioned adversarial testing across authentication and session states.
Consumer subscription application, desktop and mobile
Live Use · Greenfield · Application Software · System Owner
Took a web prototype to a code-signed, notarized desktop build and a mobile test-flight pipeline, with production payments activated. Chose a lightweight edge control plane by arguing from a no-retention requirement rather than from vendor preference. Ran three owned workstreams in parallel, each with its own session and written goal document.
Video content production pipeline
Ongoing Operation · Greenfield · Workflow Integration · System Owner
Built a repeatable pipeline from script through editorial gates, programmatic motion graphics, render, thumbnail and cross-platform promotion. Accepted final renders only against decode verification, frame counts, duration and codec assertions and content hashes. Split lead and worker sessions explicitly and gated expensive steps behind named pass/fail verdicts.
03 / Activity coverage
What the retained history can show.
559Interactive sessions
21Scheduled runs
245Active days
24Peak concurrent
The one change the same-source record supports is the appearance of scheduled, unattended runs from 20 Jul 2026 onward. Everything else observable in the window is too short to call a trend.
04 / Interaction profile
Opens with the outcome and lets the agent propose structure, supplies heavy external context up front, then becomes sharply prescriptive once there is something to react to. Corrections are located precisely; gates are placed where an action is expensive or irreversible.
Task framing
★★★
Outcome first, structure delegated
Context strategy
★★★
External material front-loaded, durable context written down
Direction and exploration
★★
Deliberate switching between the two
Correction and recovery
★★★
Names the mismatch and its location
Feedback and confirmation
★★
Gated where the action costs
05 / How you work
Build, review, operate, and maintain.
Build Review Operate
602 Authored Pull Requests Against 6 Reviewed For Others, Alongside Ongoing Operation Delivery States On Three Of Eight Arcs.
Weighted to building and operating; reviewing means judging agent output rather than other people's code.
Agent Collaboration
Three Parallel Workstreams Each With Their Own Session And Goal Document; A Median Of 9 Interactive Sessions Per Active Day And 33 Of 45 Covered Days Carrying Two Or More; No Hand Written Code In 3,419 Captured Messages.
Decomposes into owned workstreams run in parallel, then integrates and judges the combined result himself.
Verification Discipline
Second Model Plan Review, Read Only Auditor Agents, Decode And Hash Verification Of Renders, Adversarial Testing Across Authentication States — But No Test Suite Or Continuous Integration Pipeline Authored Or Run By Him.
Independent checks recur across media, interface and architecture work; the gap is an automated safety net.
Maintenance Orientation
27 Consecutive Months Of Authored Commits, 10 Repositories Worked Beyond Six Months And 5 Beyond A Year; Against That, A Scheduled Job'S Failure Was Noticed Only When An Expected Email Did Not Arrive.
Genuinely long-lived stewardship, discovered reactively rather than through instrumentation.
Favored tools
Claude CodeCodex CLIBrowser and computer useCloudflare Pages, D1 and WorkersVercelGitHub and gh CLIMail, chat and ticketing connectorsMedia toolchain
Frameworks and practices
TypeScriptNext.jsMulti-agent orchestrationCloudflare Pages, D1, KV and WorkersVisual QAAgent briefing and promptingSystems integrationScheduled automationReactSwift and native Apple platformsSQL and schema workPython
06 / Agent toolkit
The reusable operating layer.
Shell
Habitual
Present in nearly every substantive session on both tools and the dominant category by a wide margin.
Version control
Habitual
Branching, committing and pull-request work recur across projects, including ownership transfers and worktree-based parallel development.
Search and research
Habitual
Documentation lookups, competitor and reference research, and parallel research subagents feeding synthesis.
Deploy and infrastructure
Habitual
Edge and serverless deploys, domain and DNS work, scheduled automations and signing pipelines.
Browser and computer use
Habitual
Live interface review with element-level annotation, authenticated portal work and product-demo capture.
Subagents
Recurring
Parallel research and review fan-out, with single sessions reaching dozens of spawned workers.
Code editing
Recurring
Concentrated in Claude sessions; Codex applies most changes through shell patches, so this understates edit volume there.
Filesystem
Recurring
Reading specifications, playbooks and prior artifacts before work begins.
Data and database
Situational
Schema migration and query work on edge databases, concentrated in two projects.
Connectors
Situational
Mail, chat and meeting-transcript connectors used occasionally; most integration work runs through command-line tools instead.
Reusable assets
Nightly retrospective jobBounded reviewer briefModel routing policyVideo scripting, editing and promotion skillsScheduled automationsBrowser control skillMail drafting skillMarkdown review tool
Both tools run with broad standing permission that the host environment supplies rather than the person selecting per session. Against that backdrop the meaningful signal is hand-built narrowing: reviewer agents explicitly forbidden to edit files, worker briefs required to verify the repository root and fail closed on mismatch, no-spend and no-commitment limits attached to unattended automations, and a nightly job deliberately denied shell access.
07 / Subject matter
Where operating context repeats.
Across every domain the same first move recurs: decide what will count as correct — the metric, the constraint, the privacy boundary, the acceptance check — and only then delegate the building. What transfers is not a technology but that ordering.
Professional-services delivery operations
Substantial
Defined coverage vocabulary and the staffing-cost model, then kept fixing the dashboards that reported against them across five weeks. knowing which client relationships are actually covered · preparing a reviewer for a client conversation without manual assembly · attributing delivery cost to an account
Agent operating models
Substantial
Authored and revised a routing policy, standardized a fail-closed worker brief reused across three projects, and runs a nightly job that rewrites his own operating rules. deciding which model handles which class of work · bounding what an autonomous worker may touch · retaining what was learned between sessions
Services pricing and ROI modelling
Some
Negotiated fee tiers and ownership terms directly and built a calculator constrained so it cannot overstate the case. structuring a fee so both sides can defend it · showing payback without overstating it · deciding what to build rather than buy under delivery-capacity limits
Pipeline data integrity
Some
Identified a free-text field conflating two concepts and directed a separate analytical view, and traced a stale-data defect to inbound sync freshness. finding where one field carries two meanings · tracing stale downstream data to its inbound sync · deciding what an export is allowed to contain
Content production operations
Some
Accepted renders only against decode checks, frame counts and hashes, and folded each cycle's lessons back into reusable skill documents. making editorial quality repeatable rather than personal · gating expensive production steps · keeping claims current as the subject moves
Assessment and evidence systems
Some
Required older profile records to keep working unchanged as the schema advanced, and rejected a ranking iteration for adding complexity rather than removing it. deciding what counts as proof of work · keeping stored records valid as the format changes · ranking without penalising missing evidence
08 / Project fit
Staff against evidence, not labels.
Good fit
Standing up an internal operations system from scattered sources and owning it after launch
Turning a fuzzy commercial question into a defensible metric, pricing model or calculator
Designing how a team should run coding agents — routing, guardrails and reusable context
Taking a greenfield product from idea to a signed, paid, publicly reachable release
Rescuing a data-trust problem in a CRM or reporting pipeline where the definition is the real bug
Building a production pipeline with machine-checkable acceptance in a domain usually run on taste
Bring a specialist
Anything depending on a real test and continuous-integration culture — he directs testing but does not build the safety net
Security-critical work: caution and escalation are evident, but an exposed-key concern was once traded for speed
Database and schema design at scale — the constraints are his, the modelling is delegated
Accessibility-led interface work, which is absent from the evidence entirely
Reliability and observability engineering, where problems currently surface reactively
Outside current evidence
Hand-written implementation subject to code review — no self-authored code appears in any captured session
Machine-learning research, training, fine-tuning or evaluation
Reviewing and mentoring on a shared codebase he does not own, with only 6 reviewed pull requests against 602 authored
Performance and scaling engineering
Mobile-native development beyond a single test-flight release pipeline
Evidence that would improve this profile
Verification scales beyond personal review
Independent Checking Is The Strongest Pattern In This Profile, But It Currently Runs Through His Own Attention Rather Than Through Infrastructure, Which Limits How Far It Transfers To A Team.
A test suite or continuous-integration pipeline set up and depended on across at least two projects, with a failure it actually caught.
Judgment applies to code he did not commission
Only 6 Reviewed Pull Requests Appear Against 602 Authored, So Reviewer Capability Outside His Own Output Is Unestablished.
Sustained pull-request review on another person's work in a shared repository.
Depth is sustained, not just dense
Ten Of Fourteen Badges Are Held At Two Stars By The Readable Window Rather Than By The Evidence Quality, So Longer Retention Would Change The Proof Strength Directly.
Transcript-grade session history older than eight weeks, or repeat work on the same system separated by 90 or more days.
09 / Sources and limits
What this submission contains.
claude
196 Sessions
2026-06-11 → 2026-07-24
codex
374 Sessions
2025-08-19 → 2026-07-24
cowork
10 Sessions
2026-01-26 → 2026-06-10
claude-prompt-log
19422 Sessions
2025-09-28 → 2026-07-24
gemini-cli
39 Sessions
2025-06-28 → 2025-12-12
cursor
231 Sessions
2025-01-12 → 2026-06-08
Known limits
Claude transcripts cover only 11 Jun to 24 Jul 2026 (196 sessions, 44 days) because of roughly six-week local retention; Codex transcripts span 19 Aug 2025 to 24 Jul 2026 (374 sessions) but cluster in two disconnected periods with an unreadable gap between them. Depth claims are bounded by those windows.
The prompt log adds 19,422 prompts across 245 active days and 11 consecutive months from 28 Sep 2025. It carries no work product, so it supports cadence, project span and breadth only, and five months of the timeline rest on it alone.
Zero unknown-project sessions in the transcript sources; 270 records from other local tools (39 Gemini CLI sessions, 231 Cursor workspaces) carry timestamps without readable content and support no depth claim.
Delegated cloud sessions are metadata only — title and opening request — and cannot evidence implementation.
No hand-written code by the profile owner appears anywhere in 3,419 captured user messages. Every technical capability described here is directed and reviewed work, which is a genuine mode of working but distinct from implementing.
GitHub is limited to what the authenticated account can reach: unlinked commit emails, other git hosts, deleted repositories and inaccessible organizations are invisible, and 121 of 127 repositories are private. Indexed authored commits (5,549) and contribution-graph qualifying commits (105) diverge by roughly 53x for that reason and must not be merged.
Career history was not included by choice, so chronology before May 2024 is unestablished by any source.
Sessions covering personal financial, health, legal and family matters were excluded entirely during extraction. This removes real technical evidence — one excluded project was among the highest-volume in the record — and is a gap in what was read rather than in demonstrated ability.