›Hermes · Workforce Performance
Orchestration
Five agents. One hundred and sixty-nine skills. Real-time performance and learning telemetry.
Last skill update:14 min ago·Hermes runtimehealthy·Next learning cycle:04:00 GST tomorrow
Agents improve continuously. This screen shows where each one is strong, where each one is learning, and what they'll be capable of next month.
System accuracy · rolling 30d
0.0%
across 14,772 AI actions
▲ 1.8% MoM
Human corrections · this month
0
every correction trains a skill
▼ 31% MoM · fewer corrections = better performance
Skills self-improved · 30d
0
this month · across 5 agents
+12 new skills minted
Mean time to skill maturity
0 days
from skill v1 to production-grade
▼ 3 days vs Q1 · system is learning faster
Hero · The workforce
Agent performance cards
Each card is a panel — like a Bloomberg Terminal on a different security.
WT
Watson
Email & WhatsApp Employee
This month
Conversations handled12,847
Messages sent41,203
Languages active11
Avg first response34 sec
Avg conversation rating4.8 / 5.0
Accuracy by task
Tour booking99.2%
Fee questions98.7%
Document checklist99.8%
Curriculum queries96.4%
SEND / Inclusion queries94.1%
Bilingual switch (AR↔EN)97.8%
Skills
68
Total
4
Improving
62
Stable
2
New / wk
Most-improved skill · last 30 days
admissions.draft_reply.urdu
v2.1 → v2.4 · accuracy 91% → 97% · trained on 47 cases
Mostly corrections from: Hassan Mubarak
Next to mature
admissions.draft_reply.mandarin · v1.6 → v2.0 ETA Jun 8
LX
Lexa
Voice Employee · Inbound + Outbound
This month
Calls handled1,847
Avg call duration4:18
Languages active9
Handoff to human1.2%
First-call resolution88.4%
Accuracy by task
Tour scheduling99.4%
Fee acknowledgement98.1%
Inbound inquiry triage96.8%
Re-engagement (Nova handoff)94.7%
Complaint detection91.2%
Skills
42
Total
3
Improving
37
Stable
2
New / wk
Most-improved skill · last 30 days
voice.tone.warm_emirati_arabic
v1.4 → v1.9 · acceptance 84% → 96% · trained on 22 calls
Mostly corrections from: Hassan Mubarak
Next to mature
voice.tone.warm_hindi_devotional · v1.2 → v1.6 ETA Jun 12
SH
Sherlock
Research & Pattern Detection
This month
Queries answered4,212
Patterns detected (Intel)18
Cross-source joins14,890
Data sources spanned14
Avg time-to-answer3.4 sec
Accuracy by task
Data retrieval (single src)99.7%
Multi-source join97.2%
Pattern detection81.6%
Anomaly flagging89.4%
Citation quality99.1%
Skills
31
Total
2
Improving
27
Stable
2
New / wk
Most-improved skill · last 30 days
research.cross_school_pattern
v1.3 → v1.7 · false-positive rate 18% → 8%
Mostly corrections from: Dr. Neil Hopkin
Next to mature
research.safeguarding_pattern_detector · v2.1 → v2.4
NV
Nova
Outbound Re-engagement
This month
Reach-outs sent340
Responses received94
Re-engagement rate27.6%
Tour bookings from re-engaged14
Languages used8
Accuracy by task
Tone match (soft consultative)96.8%
Timing (avoid weekends / late)100.0%
Language match98.4%
Personalisation depth88.2%
Skills
18
Total
2
Improving
15
Stable
1
New / wk
Most-improved skill · last 30 days
outbound.reengagement_warmth
v1.2 → v1.6 · response rate 18% → 28%
Pattern learned: families with elder siblings respond 23% better to "we missed seeing your name on the renewal" framing vs generic "checking in".
RY
Ray
Orchestrator
This month
Decisions made47,212
Routing accuracy99.2%
Escalations to humans519
Failed approval-checks0 ✓
Cross-agent handoffs8,140
Accuracy by task
Language routing99.8%
Brand context detection99.4%
Priority scoring96.1%
Escalation decisioning97.8%
Conflict resolution94.2%
Skills
10
Total
1
Improving
9
Stable
0
New / wk
Most-improved skill · last 30 days
orchestration.escalation_threshold
Tuned by 47 human override events
Now matches David Williams' personal preference 94% of the time.
Policy rules in effect
14
Violations this month
0 ✓
Telemetry · Versioned skills
Skill evolution — last 90 days
Hover any version marker to see what trained it.
Skill evolution — last 90 days
in productionversion release
~ 90 days agotoday
admissions.draft_reply.urdu
Watson
v1.8 · trained on corrections · acc lift +2%
v1.8v2.0 · trained on corrections · acc lift +4%
v2.0v2.1 · trained on corrections · acc lift +6%
v2.1v2.4 · trained on corrections · acc lift +8%
v2.4admissions.draft_reply.arabic
Watson
v2.6 · trained on corrections · acc lift +2%
v2.6v2.7 · trained on corrections · acc lift +4%
v2.7v2.9 · trained on corrections · acc lift +6%
v2.9v3.1 · trained on corrections · acc lift +8%
v3.1v3.2 · trained on corrections · acc lift +2%
v3.2voice.tone.warm_emirati_arabic
Lexa
v1.0 · trained on corrections · acc lift +2%
v1.0v1.4 · trained on corrections · acc lift +4%
v1.4v1.7 · trained on corrections · acc lift +6%
v1.7v1.9 · trained on corrections · acc lift +8%
v1.9research.cross_school_pattern
Sherlock
v1.0 · trained on corrections · acc lift +2%
v1.0v1.3 · trained on corrections · acc lift +4%
v1.3v1.5 · trained on corrections · acc lift +6%
v1.5v1.7 · trained on corrections · acc lift +8%
v1.7outbound.reengagement_warmth
Nova
v1.0 · trained on corrections · acc lift +2%
v1.0v1.2 · trained on corrections · acc lift +4%
v1.2v1.6 · trained on corrections · acc lift +6%
v1.6orchestration.escalation_threshold
Ray
v3.1 · trained on corrections · acc lift +2%
v3.1v3.2 · trained on corrections · acc lift +4%
v3.2v3.4 · trained on corrections · acc lift +6%
v3.4v3.7 · trained on corrections · acc lift +8%
v3.7report.eoy_year6.uk_english
Watson
v1.0 · trained on corrections · acc lift +2%
v1.0v1.2 · trained on corrections · acc lift +4%
v1.2safeguarding.pattern_detector
Sherlock
v2.0 · trained on corrections · acc lift +2%
v2.0v2.1 · trained on corrections · acc lift +4%
v2.1v2.2 · trained on corrections · acc lift +6%
v2.2v2.4 · trained on corrections · acc lift +8%
v2.4fee_reminder.tone.soft
Watson
v1.4 · trained on corrections · acc lift +2%
v1.4v1.6 · trained on corrections · acc lift +4%
v1.6v1.8 · trained on corrections · acc lift +6%
v1.8v2.0 · trained on corrections · acc lift +8%
v2.0fee_reminder.tone.firm_polite
Watson
v0.1 · trained on corrections · acc lift +2%
v0.1Each new version is the result of human corrections and direct feedback. Hermes runtime trains weekly. No skill is released without 7-day shadow testing.
Loop · Human teaches, system learns
Human-in-the-loop quality
The relationship between corrections and accuracy — and the people whose feedback shapes the agents.
Correction volume vs accuracy · 90 days
Daily corrections (left axis)Daily AI accuracy (right axis)
As corrections decrease, accuracy increases. The system is learning your standards.
Top correctors · who trains the most
Hassan Mubarak
multilingual
142
Sarah Pereira
admissions
98
Dr. Neil Hopkin
curriculum
64
Devika Singh Mankani
SEND
47
David Williams
escalations
38
Nicholas Rickford
Sunmarke voice
34
Each correction is a teaching moment. The team that corrects most shapes the agent that everyone uses.
System health & reliability
Hermes runtime
Healthy
99.97% uptime (90d)
Inference latency · p50
142ms
target < 500ms
Inference latency · p99
814ms
target < 2000ms
Knowledge base sync lag
4.2 sec
target < 60 sec
Skill registry consistency
Verified
last check 09:00 GST
Audit log integrity
Sealed
hash chain verified
Incident history · last 90 days
- •1 minor degradation · Apr 12 · 14m (Facilio vendor-side, not Co-pilot)
- •0 customer-impacting incidents
- •0 data integrity events
- •0 security events
Coming next week
4 releases scheduled
🆕
admissions.draft_reply.mandarin v2.0
Graduates from human-review to auto-send after 2-week shadow period.
ETA Tuesday · Jun 3
🆕
research.safeguarding_pattern_detector v2.4
Reduced false positives by 38%. Reviewed by Dr. Hopkin.
ETA Wednesday · Jun 4
🆕
outbound.reengagement_warmth v1.7
New variant for "elder sibling recent leaver" segment.
ETA Friday · Jun 6
🆕
fee_reminder.tone.firm_polite v0.5
New branched tone for second-attempt reminders. Pending Sunita's voice review.
ETA next week
Every conversation, every correction, every approval shapes the agents. Co-pilot is not deployed once and left alone — it is a living system. As your team's standards rise, the agents rise to meet them. As your business changes, they adapt. The platform you pay for in June will not be the same platform you have by December. It will be measurably better, in directions your team will have chosen.
Orchestration settings · click any to edit · changes audit-logged
Hermes · standing by · next cycle 04:00