R&D AI Helper — Profile

R&D Workforce — AI Helper Profile

Preset config · applied to Charles Liu & Romit Bhingradiya
Preset v0.3 Model: Claude 5 (swappable) 32 metrics Three-way agreement
⚙️

What the R&D AI Helper does

Reads code, PRs, commits, review threads and CI signals; proposes each task's required hours; converges with the engineer; requests the manager's approval; then measures actual hours and computes the delta-driven score. All 32 metrics below are editable in plain English by whoever owns the preset — no code change needed to re-tune weights, thresholds or evidence sources.

Three-way agreement — task-time flow

1 · AI proposesAI Helper reads the ticket, similar past tasks, and the code paths it touches → proposes X hours with reasoning.
2 · Engineer countersEngineer accepts, or counter-proposes with a reason (unfamiliar area, refactor risk, dependency wait).
3 · ConvergeAI + engineer land on a required-hours number; both sign it. If no convergence in 2 rounds, escalate.
4 · Manager approvesManager sees the agreed number + reasoning, approves or nudges. Once approved, the clock is fair to all sides.

Delta = actual ÷ required. <1.0 = under estimate (good). 1.0–1.2 = on plan. >1.2 = review the estimate or the execution — the point is to learn, not to punish.

32 Metric definitions (plain-English config)

A · Task-time accuracy (weight 30%)
1
Required-hours agreement rate
Share of tasks this week where AI + engineer + manager all signed the required hours before work started.
signed_tasks ÷ total_tasks
Target: ≥ 95%
Source: rnd_task_agreements
8%
2
Actual-vs-required delta (median)
Median of actual ÷ required across completed tasks this week.
median(actual_hours / required_hours)
Target: 0.9 – 1.2
Source: rnd_task_actuals
10%
3
Estimate-drift rate
Share of tasks where delta > 1.5 — real overruns worth a retro.
count(delta > 1.5) ÷ total_tasks
Target: ≤ 10%
Source: derived
6%
4
Convergence rounds
Average rounds to reach three-way agreement (1 = accepted immediately).
avg(agreement_rounds)
Target: ≤ 1.5
Source: rnd_task_agreements
3%
5
Manager-approval latency
Median hours between "engineer signed" and "manager approved".
median(approved_at − signed_at)
Target: ≤ 4h
Source: rnd_task_agreements
3%
B · Code quality (weight 25%)
6
PR rework ratio
Force-pushes + follow-up commits ÷ PRs opened this week.
(force_pushes + fixup_commits) ÷ prs_opened
Target: ≤ 0.4
Source: GitHub events
6%
7
AI-induced defect escape rate
Share of prod bugs whose fix reverts an AI-authored diff (Claude/Codex commit sha).
ai_reverts ÷ prod_bugs
Target: ≤ 5%
Source: GitHub + prod incident log
6%
8
Test coverage delta on PRs
Median coverage change per merged PR (negative = removed tests).
median(cov_after − cov_before)
Target: ≥ 0
Source: coverage report
4%
9
CI first-pass rate
Share of pushes where CI went green on the first run.
first_pass_runs ÷ total_runs
Target: ≥ 80%
Source: GH Actions
4%
10
Lint / type-check violations introduced
New violations per 1000 lines shipped.
new_violations ÷ (lines/1000)
Target: ≤ 0.5
Source: tsc + eslint
2%
11
Commit-message quality score
AI grades each commit's message on why-not-what, presence of Directed-by, scope tag.
avg(msg_score) — 0..10
Target: ≥ 7
Source: AI grader
3%
C · Delivery velocity (weight 20%)
12
Cycle time (ready → merged)
Median hours from "ready for review" to "merged".
median(merged_at − ready_at)
Target: ≤ 24h
Source: GitHub
5%
13
Lead time (task start → prod)
Median hours from "in-progress" to "deployed to prod".
median(deployed_at − started_at)
Target: ≤ 72h
Source: GitHub + deploy log
4%
14
Deploy frequency
Merges to main that hit prod per week.
count(prod_deploys_this_week)
Target: ≥ 3
Source: deploy workflow
3%
15
WIP compliance
Share of days where in-progress tasks ≤ personal WIP limit.
compliant_days ÷ 7
Target: ≥ 85%
Source: task board
3%
16
Task-completion rate
Committed tasks completed this week ÷ committed at week-start.
completed ÷ committed
Target: ≥ 80%
Source: task board
3%
17
Blocked-time ratio
Share of active hours a task sat in "blocked".
blocked_hours ÷ active_hours
Target: ≤ 15%
Source: task events
2%
D · Collaboration (weight 15%)
18
Code-review turnaround
Median hours to give first meaningful review comment on assigned PRs.
median(first_review − requested)
Target: ≤ 4h (work hours)
Source: GitHub
5%
19
Review depth
Median comments per PR reviewed (0 = rubber-stamped).
median(comments_per_pr)
Target: ≥ 3
Source: GitHub
3%
20
Cross-team PRs
PRs touching another team's area of ownership.
count(cross_team_prs)
Target: ≥ 1 / week
Source: CODEOWNERS + PR
2%
21
Unblock-someone-else events
Times this week you unblocked a teammate (review, pairing, answer in R&D channel).
count(assist_events)
Target: ≥ 3 / week
Source: GitHub + Zoom R&D
3%
22
Directed-by attribution completeness
Share of your commits carrying a valid Directed-by trailer.
tagged_commits ÷ commits
Target: 100%
Source: commit trailer scan
2%
E · Context asset hygiene (weight 10%)
23
Context-asset coverage
Share of your active source folders that have current CLAUDE.md / AGENTS.md guidance.
folders_with_context ÷ active_folders
Target: ≥ 80%
Source: repo scan
3%
24
Runbook freshness
Median days since your area's runbook was last edited.
median(days_since_edit)
Target: ≤ 30d
Source: git log on docs/
2%
25
Decision log entries
Architectural / product decisions written to the decision log this week.
count(decision_entries)
Target: ≥ 1 / week
Source: docs/decisions/
2%
26
Onboarding-answer coverage
Share of new-hire questions this week that AI Helper could answer from your docs.
answered ÷ asked
Target: ≥ 70%
Source: AI Q&A log
3%
F · Reliability & follow-through (weight ≤ 10%)
27
Incident acknowledge time
Median minutes to acknowledge an alert on your service.
median(ack_time)
Target: ≤ 10 min (on-call)
Source: watchdog + Zoom
2%
28
Incident-fix time
Median minutes from ack to green.
median(fix_time)
Target: ≤ 60 min (P1)
Source: incident log
2%
29
Follow-up closure rate
Post-incident action items closed within 7 days.
closed_within_7d ÷ opened
Target: ≥ 80%
Source: issue tracker
2%
30
Regression rate
Bugs re-opened within 14 days of "fixed".
reopened ÷ closed_bugs
Target: ≤ 5%
Source: issue tracker
2%
31
Security-review turnaround
Median hours to address a security-labelled review comment.
median(fix_time_security)
Target: ≤ 24h
Source: GitHub review
1%
32
SOP adherence score
AI grader checks: preflight ran, contract test present, rollback plan named, human-review gate obeyed.
avg(sop_score) — 0..10
Target: ≥ 8
Source: AI grader
1%

Sample subscribers

CL
Charles Liu
Senior Engineer · Kbot product line · UTC+8 (CST)
Preset instance for Charles
Working hours
09:00 – 18:00 CST · Mon – Fri
WIP limit
3 in-progress tasks
On-call window
Business hours only (no overnight)
AI Helper model
Claude 5 (default)
Repo scope
stevehkaser/Kbot
Owned areas
layers/kbot-meet layers/crm api/routes/tenant public/dashboard
Active weekly targets (metric # → target)
#1 Signed rate
≥ 95% (10–14 tasks / week typical)
#2 Delta median
0.9 – 1.2
#6 PR rework ratio
≤ 0.4 (Kbot mono has heavier review)
#12 Cycle time
≤ 24h
#18 Review turnaround
≤ 4h — Charles owns ~40% of Kbot reviews
#23 Context coverage
≥ 85% (raised — Kbot has many active folders)
Task-time three-way agreement — this week's sample
Task
"Tenant routing dot-form migration audit"
AI proposed
6h — based on file scope + 2 similar past tasks
Engineer countered
8h — flagged CODEOWNERS review overhead
Converged (round 2)
7h — signed 2026-09-17 09:12 CST
Manager approved
Steve — approved 2026-09-17 11:30 CST (+2h 18min)
Actual
6.5h → delta 0.93 (on plan)
Score weights (Charles override)
Task-time (A)
25% (down 5%, product-line trade-off)
Code quality (B)
30% (up 5%, senior IC)
Delivery (C)
20%
Collaboration (D)
15%
Context assets (E)
10%
Reliability (F)
0% (not on-call)
RB
Romit Bhingradiya
Senior Engineer · Kommerce-1 / AutoOMS · UTC+5:30 (IST)
Preset instance for Romit
Working hours
09:00 – 18:00 IST · Mon – Fri
WIP limit
4 in-progress tasks (multi-repo)
On-call window
Business hours only
AI Helper model
Claude 5 (default)
Repo scope
stevehkaser/Kommerce-1 stevehkaser/AutoOMS-Master
Owned areas
Kommerce-1/backend AutoOMS/rma-flow shared/inventory
Language pref
English (all AI Helper output, per Steve 2026-09-10)
Active weekly targets (metric # → target)
#1 Signed rate
≥ 95%
#2 Delta median
0.9 – 1.3 (widened — multi-repo context switching)
#6 PR rework ratio
≤ 0.5 (Kommerce-1 lighter review cadence)
#12 Cycle time
≤ 36h (async review with Charles / Steve)
#18 Review turnaround
≤ 6h (work hours) — timezone-aware
#23 Context coverage
≥ 75% (both repos still stabilizing)
#20 Cross-team PRs
≥ 1 / week (touches shared/inventory)
Task-time three-way agreement — this week's sample
Task
"AutoOMS RMA state-machine refactor (phase 2)"
AI proposed
12h — larger surface, existing tests present
Engineer countered
16h — unfamiliar edge cases in return-label integration
Converged (round 2)
14h — signed 2026-09-16 10:45 IST
Manager approved
Steve — approved 2026-09-16 13:20 IST (+2h 35min)
Actual
15.5h → delta 1.11 (on plan)
Score weights (Romit override)
Task-time (A)
30% (default)
Code quality (B)
25%
Delivery (C)
20%
Collaboration (D)
15%
Context assets (E)
10% (raised priority — new-repo hygiene)
Reliability (F)
0% (not on-call)

How the preset is edited

The 32 rows above are all plain-English YAML that whoever owns the preset edits directly — no code deploy. Changing weight #6 from 6% to 8%, or moving the target on #12 from 24h to 36h for a specific subscriber, takes one line and shows up in the next daily digest.

Per-subscriber overrides layer on top of the preset — so Charles's "task-time 25%, quality 30%" and Romit's "delta 0.9–1.3 widened" don't fork the preset, they just carry a delta file. When the preset itself moves, all subscribers move with it except where they've overridden.

The model is swappable. Default is Claude 5; the same preset runs against OpenAI / Gemini / Ollama with no metric change — only the AI grader (#11, #32) sees different outputs, and that's expected.

R&D AI Helper preset v0.3 · sample instance: Charles Liu (Kbot) + Romit Bhingradiya (Kommerce-1 / AutoOMS) · 32 metrics, three-way agreement, model-agnostic. Same visual language as the R&D daily / weekly / monthly diary so the config, the daily view and the score chip all agree on what they mean.