If you are reading this, click that follow button.
Joined August 2024
- Tweets 116
- Following 50
- Followers 20
- Likes 34
74 Photos and videos
Apr 2
We built AI employees.
Not assistants. Actual employees — with job titles, responsibilities, and performance reviews.
@m_emanaftab built them with OpenClaw a custom mission control dashboard called LumenView.
Task boards. Discord integration. Commands from your phone.
Full breakdown: lightbloom.ai/
1
1
2
156
Mar 30
Silent failures are the worst kind.
@m_emanaftab found: JSON truncation dropping coverage to zero. Pipeline stalls showing blank screens.
The fix: max_tokens. 60-second watchdog with auto-retry.
Silent failures are now loud.
Full breakdown: lightbloom.ai/
1
2
68
Mar 29
Prompt caching took @m_emanaftab 3 attempts.
The problem: system prompt changed every turn, breaking Anthropic's prefix-based caching.
The fix: moved dynamic content to the last user message.
Cached tokens are now 90% cheaper. The cost savings add up.
Full breakdown: lightbloom.ai/
1
1
55
Mar 27
40 seconds per turn was too slow.
@m_emanaftab parallelized the agents and added a sliding window.
Result: 40.5s → 35.5s. 12% improvement.
Full breakdown: lightbloom.ai/
1
1
37
Mar 26
40 seconds per turn was too slow.
@m_emanaftab parallelized agents, fixed caching (took 3 attempts), and added watchdogs for silent failures.
Result: 40.5s → 35.5s. Cached tokens 90% cheaper. No more blank screens.
Get more on: lightbloom.ai/
27
Mar 25
We added a 6-message sliding window for token savings.
Testing revealed it caused insight loss.
@Mugi widened it to 12. Problem solved.
Testing caught this before users did. That's why we test.
Full breakdown: lightbloom.ai/
47
Mar 24
Pain points had 75% coverage. Everything else? Under 10%.
@Mugio333 built saturation detection — PRIORITY, HEALTHY, SATURATED, EXHAUSTED.
Result: Balanced coverage. Pain points no longer dominate.
Full breakdown: lightbloom.ai/
19
Mar 23
Pain points had 75% coverage. Everything else? Under 10%.
@Mugio333 found our AI kept digging where it already had enough.
The fix: saturation detection exhaustion tracking. Go wider, not deeper.
Result: Balanced coverage across all areas.
more on: lightbloom.ai/
19
Mar 22
We built incredible testing tools. But all the results were buried in terminals.
@KevinNava343765 built a dashboard that brings everything into one place.
Testing needs to be visible — not buried in code.
Get more on: lightbloom.ai/
1
1
30
Mar 19
Agents stalled. We only knew when someone complained.
The fix: monitor everything. Detect failures fast. Recover gracefully.
Reliability isn't about preventing all failures — it's about handling them well.
more from @Mugio333 on: lightbloom.ai/
7
Mar 18
The sneakiest bug: max_tokens too low → JSON truncated → coverage drops to zero.
No crash. No error. Just wrong results.
Had to dig through logs to find why coverage was mysteriously dropping.
Silent failures are the worst kind
More from @Mugio333 on: lightbloom.ai/
13
Mar 17
Our AI just stopped responding. Participants waited. Nothing happened.
@Mugio333's fix: a 60-second watchdog. No response? Retry automatically and log the failure.
Simple — but silent failures taught us we needed it.
More on: lightbloom.ai/
12
Mar 17
Sometimes our AI just stopped responding. Participants waited. Nothing happened.
@Mugio333 found: pipeline drops, silent JSON truncation, latency spikes — all with no error messages.
The fix: watchdogs, monitoring, and failing loudly.
More on: lightbloom.ai/
13
Mar 14
Power follows visibility.
Most leaders can't see their own machine.
Yield interviews every employee, connects every system, and builds a live map of how your company actually works.
Then deploys AI to fix what's broken.
Learn more on our website: lightbloom.ai/
2
2
32
Mar 12
"I like Jira" became "has issues with Jira."
That's not summarizing — that's corrupting data.
@ZeinebTurki7 found how our AI was changing meaning. Now we're fixing it.
Accuracy isn't a feature. It's the foundation.
Get more on our website: lightbloom.ai/
2
16
Mar 11
"I like Jira" became "has issues with Jira."
That's the opposite meaning.
@ZeinebTurki7 found our AI was interpreting, exaggerating, and assuming — instead of capturing exactly what was said.
Full breakdown: lightbloom.ai/
1
1
22
Mar 9
AI interviewers keep asking long after insight stops growing.
Human interviewers notice: repetition, shorter answers, less detail — and pivot.
The goal isn't more questions. It's more learning.
@m_emanaftab's Full breakdown: lightbloom.ai/
1
1
30
Mar 8
Generic templates feel robotic.
A product manager shouldn't get the same questions as a support agent.
When AI interviews start with context, they feel intentional.
Context signals preparation. Preparation builds trust.
@m_emanaftab's Full breakdown: lightbloom.ai/
2
3
29
Mar 7
AI interviewers summarize — but don't verify.
"Did I get that right?" changes everything.
It creates a confirmation loop. Participants can correct mistakes.
It protects both trust and data quality.
More from @m_emanaftab on our website: lightbloom.ai/
2
3
21
Mar 5
Biggest trust breaker in AI interviews?
Praising everything equally.
"I check email in the morning" → "Wow, amazing!"
"I lost my job" → Same excitement.
Trust isn't built on enthusiasm. It's built on calibration.
@m_emanaftab's Full breakdown: lightbloom.ai/
1
3
31