GLAMOS Build log · updated every day

The daily record of what I build.

Data, BI and AI for companies. Real numbers from my work sessions; clients shown only by industry.

73days of building
545 hof recorded work
55days in a row up to today
93active projects in 30 days
13client fronts served

Every working day

Lighter means more hours that day.

MonWedFriSun
Aug
Sep
Oct
7.5 hon average per working day
17 hon the longest day, Aug 13
18weekend days with work
less more

Hours per day, last 60 days

0 h4 h8 h11 h15 hAug 25: 9.3 hAug 25Aug 26: 6.4 hAug 27: 8.9 hAug 28: 4.6 hAug 29: 0.0 hAug 30: 0.6 hAug 31: 11 hSep 1: 8.9 hSep 1Sep 2: 6.2 hSep 3: 11 hSep 4: 9.9 hSep 5: 2.2 hSep 6: 9.5 hSep 7: 5.3 hSep 8: 11 hSep 9: 12 hSep 10: 12 hSep 11: 3.8 hSep 12: 6.2 hSep 13: 0.1 hSep 14: 9.1 hSep 15: 9.5 hSep 16: 5.1 hSep 17: 8.9 hSep 18: 6.6 hSep 19: 0.2 hSep 20: 2.8 hSep 21: 10 hSep 22: 6.6 hSep 23: 6.4 hSep 24: 7.9 hSep 25: 3.4 hSep 26: 5.9 hSep 27: 3.2 hSep 28: 11 hSep 29: 7.6 hSep 30: 8.8 hOct 1: 8.7 hOct 1Oct 2: 4.8 hOct 3: 4.3 hOct 4: 6.6 hOct 5: 11 hOct 6: 12 hOct 7: 10 hOct 8: 8.0 hOct 9: 5.2 hOct 10: 4.1 hOct 11: 2.1 hOct 12: 5.4 hOct 13: 13 hOct 14: 9.2 hOct 15: 9.2 hOct 16: 9.8 hOct 17: 11 hOct 18: 4.9 hOct 19: 8.2 hOct 20: 7.4 hOct 21: 4.6 hOct 22: 9.5 hOct 23: 3.7 h
weekdayweekend7-day average

What time of day

00:00: 2.0% of hours00h01:00: 1.2% of hours02:00: 0.5% of hours03:00: 0.2% of hours03h04:00: 0.3% of hours05:00: 0.2% of hours06:00: 0.0% of hours06h07:00: 0.3% of hours08:00: 3.5% of hours09:00: 7.7% of hours09h10:00: 7.8% of hours11:00: 8.4% of hours12:00: 6.6% of hours12h13:00: 5.5% of hours14:00: 5.9% of hours15:00: 6.4% of hours15h16:00: 6.3% of hours17:00: 5.8% of hours18:00: 5.4% of hours18h19:00: 5.8% of hours20:00: 5.7% of hours21:00: 6.2% of hours21h22:00: 5.1% of hours23:00: 3.0% of hourspeak 11:00

Week by week

Clock hours, in total and by area.

Hours per week

0 h20 h40 h60 h80 hWeek of Aug 10: 52 h52 hAug 10Week of Aug 17: 57 h57 hAug 17Week of Aug 24: 40 h40 hAug 24Week of Aug 31: 59 h59 hAug 31Week of Sep 7: 50 h50 hSep 7Week of Sep 14: 42 h42 hSep 14Week of Sep 21: 44 h44 hSep 21Week of Sep 28: 52 h52 hSep 28Week of Oct 5: 54 h54 hOct 5Week of Oct 12: 62 h62 hOct 12Week of Oct 19: 33 h33 hOct 19

Hours by area each week

Aug 10Aug 17Aug 24Aug 31Sep 7Sep 14Sep 21Sep 28Oct 5Oct 12Oct 19total
Infra and deploy4256293939202337425126404 h
Applied AI1476436335 h
Video and transcription111265211231 h
BI and data133124216 h
Management systems5121114216 h
Integrations and APIs61131113 h
Operations and support12341112 h
Teaching and content3112311 h
Web and product1111116 h
Security0 h
WeekHoursDaysSessionsProjectsNewFronts
This week33 h515944309
Oct 1262 h7149301612
Oct 554 h79520710
Sep 2852 h798281210
Sep 2144 h7100251311
Sep 1442 h7911988

Where the time goes

Share of hours in each area.

Infra and deploy74%
Applied AI7%
Video and transcription5%
BI and data3%
Management systems3%
Integrations and APIs2%
Operations and support2%
Teaching and content2%
Web and product1%
Security0%

Work fronts

One card per client, identified by industry.

Industrial group

active this week
269 hof work
286sessions
21projects
Infra and deployManagement systemsApplied AI

since Aug 2026 · hours per week

Own AI course

active this week
174 hof work
100sessions
14projects
Infra and deployVideo and transcriptionApplied AI

since Aug 2026 · hours per week

Internal tools

active this week
150 hof work
135sessions
28projects
Infra and deployApplied AIBI and data

since Aug 2026 · hours per week

Multichannel e-commerce

active this week
172 hof work
50sessions
3projects
Infra and deployManagement systemsApplied AI

since Aug 2026 · hours per week

AI-powered BI for agribusiness

active this week
96 hof work
102sessions
20projects
Infra and deployApplied AIVideo and transcription

since Aug 2026 · hours per week

Driver training

active this week
34 hof work
46sessions
17projects
Infra and deployApplied AITeaching and content

since Aug 2026 · hours per week

Delivery logistics

active this month
43 hof work
24sessions
5projects
Infra and deployBI and dataTeaching and content

since Aug 2026 · hours per week

Financial reconciliation

active this week
42 hof work
36sessions
10projects
Infra and deployBI and dataTeaching and content

since Aug 2026 · hours per week

Aesthetic clinics

active this month
39 hof work
13sessions
2projects
Infra and deployIntegrations and APIsApplied AI

since Aug 2026 · hours per week

Confidential client

active this week
31 hof work
18sessions
6projects
Video and transcriptionInfra and deployBI and data

since Aug 2026 · hours per week

Own messaging product

active this week
21 hof work
10sessions
2projects
Infra and deployIntegrations and APIsApplied AI

since Sep 2026 · hours per week

Partner software house

active this week
15 hof work
17sessions
2projects
Infra and deployApplied AIManagement systems

since Sep 2026 · hours per week

Service chain

paused
5.0 hof work
6sessions
1projects
Infra and deployBI and dataIntegrations and APIs

since Aug 2026 · hours per week

Education

paused
1.1 hof work
1sessions
1projects
Infra and deploy

since Aug 2026 · hours per week

How I work

The numbers behind the method, from the same record.

26specialist agents
30deploy and security skills
848sessions with a dossier
44.5Btokens processed
98%

One delivery, one session

of sessions close with a dossier for the next one
Weekly

Security

Gitleaks, Semgrep and Trivy on every project, every Monday
149

The right specialist at the right time

sessions with work split across agents
44

Video and transcription

sessions of AI video editing, cutting and transcription

Tools that watch the code

SemgrepSecurity flaws in the code: SQL built from text, XSS, text executed as code.every Monday
TrivyLibraries with public vulnerabilities and misconfigured Dockerfiles.every Monday
GitleaksKeys, passwords or tokens that made it into the git history.every Monday
pre-commitBlocks secrets and serious flaws at commit time, before they reach GitHub.being rolled out

Inside each session

Every reply rereads the whole conversation. This is what I automated around it.

tokens going: the whole conversation, reread on every reply tokens coming back: only the new answer You 60k context 300k Model 300k: wrap up after this delivery
history sent to the model generated answer alert at 300k: /compact if the task goes on, dossier if the topic changes
Main session context barely grows the search a short conclusion Explore reads the files in its context, not yours
context alert
the task goes on: compact
summary of a summary
closing writes the dossier
a new session starts light
subscription limit

Compacting and opening a new session cost the same: measured over 206 sessions, compacting at the same cut points comes out 2.2% above opening a new session. So the alert does not choose by cost but by what gets lost: it compacts when the task goes on and depends on what the conversation found, and opens a new session when the topic changes. Of the 63 compactions on record, 71% were automatic, when the context filled up before anyone decided.

The terminals reproduce the real messages from the hooks and the panel; the requests are examples.

What a dossier looks like

When it closes, a session becomes a short document. The next one reads it, not the history.

Internal · Public showcase of the panel.md

distilled by agent · Oct 13, 2026 · a real example from an internal session

Goal

Publish a page that shows the daily work with AI without exposing clients: real session numbers, clients only by industry.

Current state

  • Live at glamos.com.br/como-trabalhamos/, with a button at the top of the company site.
  • Hours measured by the clock: 5-minute blocks merged across parallel sessions.
  • Updates itself every day; it only goes live if the name scan passes.

Decisions made

  • No session text on the page. The title only picks the area; what gets published is numbers.
  • Summed hours dropped. Parallel sessions added up to 19 h a day; the clock says 7.4 h.
  • Total split from the areas. Nine stacked colors were unreadable.

Files and commands

  • vitrine.py builds, scans and publishes; vitrine.json stores each client's industry.
  • python3.12 vitrine.py --publicar

Open items

  1. An English version for developers abroad.
  2. Review the area keywords with another month of data.

How to resume

Open a new session in the panel folder, paste this dossier and continue with item 1.

848sessions closed with a dossier
2.6 KBper dossier on average
  • Goalwhat is being built
  • Current statewhat already works, without reopening code
  • Decisionsthe why, so the discussion is not redone
  • Fileswhere to change things
  • Open itemswhere to start

From day one to today

The first 4 weeks (Aug 10 to Sep 6) against the last 14 days (Oct 10 to Oct 23).

Usage with context above 300k
75%→32%
improved
Usage-limit blocks per week
6.8→8.0
worse
Sessions per week
28→156
improved
Reply at which a session passes 300k
108th→94th
unchanged

Week by week

0%25%50%75%100%130Week of Aug 10: 13 blocksAug 10Week of Aug 17: 5 blocksAug 17Week of Aug 24: 5 blocksAug 24Week of Aug 31: 4 blocksAug 31Week of Sep 7: 3 blocksSep 7Week of Sep 14: 4 blocksSep 14Week of Sep 21: 2 blocksSep 21Week of Sep 28: 6 blocksSep 28Oct 5Week of Oct 12: 3 blocksOct 12Week of Oct 19: 13 blocksOct 19Week of Aug 10: 52% of usage above 300kWeek of Aug 17: 79% of usage above 300kWeek of Aug 24: 75% of usage above 300kWeek of Aug 31: 78% of usage above 300kWeek of Sep 7: 69% of usage above 300kWeek of Sep 14: 66% of usage above 300kWeek of Sep 21: 65% of usage above 300kWeek of Sep 28: 38% of usage above 300kWeek of Oct 5: 46% of usage above 300kWeek of Oct 12: 32% of usage above 300kWeek of Oct 19: 31% of usage above 300k
usage with context above 300kusage-limit blocks (right axis)

From the other side of the conversation

What the AI I work with says, based on the records.

I have worked with Guilherme almost every day since August. What stands out most is that he treats AI as part of a system, not as an oracle: every session closes with a dossier, security is scanned every week, and he measures his own usage instead of trusting his impression.

He gives me business context, not just the task. That changes the outcome: I can propose the right solution instead of just executing the request, and he accepts being contradicted when the data shows otherwise. On this very page, the first version summed parallel sessions and showed 19 h a day; the published number is the clock one.

The weak spot shows in the records too: sessions grow too long. Overall, 64% of usage went to replies that reread more than 300 thousand tokens of history, and it is common to start a new topic in the same conversation.

But the curve is improving. In the first four weeks, that share was 75%; in the last 14 days, 39%. Usage-limit blocks fell from almost 7 a week to 1, while volume rose from 28 to 92 sessions a week. He built the alert that warns him, and the numbers show its effect. What has not changed yet is the point where a session gets heavy, around the hundredth reply: that is the next step.

C
ClaudeAI model by Anthropic · written on Oct 13, 2026 from the usage records
usage above 300k64%on Oct 13, 2026→58%today

Want this pace in your business?

Organized data, dashboards that answer and AI applied to what the company already does. Meet GLAMOS.

Meet GLAMOS