A Software Engineer who strongly believes one can add immense value by knowingly implementing Software Engineering best practices, even at the age of AI.
A Software Engineer who strongly believes one can add immense value by knowingly implementing Software Engineering best practices, even at the age of AI.
Maybe DEV got inspired by my Dev user name and came up with this challenge, lol! Just kidding 🤣By the way, DEV should launch a Halloween challenge too! 🎃👻
A Software Engineer who strongly believes one can add immense value by knowingly implementing Software Engineering best practices, even at the age of AI.
Just published my submission for Week 3: SlothLog (🦥)!
It’s an offline, privacy-first developer utility that inspects local git commits, clusters them into work sessions by ticket ID, and writes clean timesheet worklogs using local Gemma 2 2B (with Sentry tracing and a Render cloud demo). No more wasting 45 minutes every Friday afternoon trying to remember what you did on Tuesday!
Hi Alessio! I read your write-up—the ticket grouping and editable hour controls are useful touches. I'm working on a nearby problem with SlothPulse: local Gemma drafts a standup from commits and an optional traceback, with Python binding each item to its source.
One failure surprised me: a summary cited the correct commit but turned ‘submission handoff’ into ‘submitted.’ Source validation passed, yet the meaning still needed correction. Have you seen Gemma 2 2B turn tentative work into a completed accomplishment, especially with short ‘wip’ commits? Curious how you handle that before copying a worklog.
Hi @tengbyte! Thanks so much for reading and for the thoughtful question! SlothPulse sounds like a fantastic concept—binding standup bullets to source commits and tracebacks is super smart.
You hit on a very real challenge: compact models like Gemma 2 2B definitely suffer from an inherent "completion bias". Because summarization fine-tuning heavily associates commit logs with deliverables, the model loves to turn tentative phrasing like "wip: auth handoff" into triumphant assertions like "Completed auth implementation."
Here’s how we approach and mitigate this in SlothLog:
Strictly Low Temperature (0.2): We keep Ollama’s inference temperature pinned at 0.2 to minimize creative extrapolation and keep Gemma tightly tethered to the actual commit verbs.
Session Clustering Context: Rather than feeding isolated commits, SlothLog clusters commits within a 90-minute activity window. When Gemma sees a sequence of commits (e.g. wip refactoring, investigating edge case, drafting tests), the chronological timeline helps anchor the reality of the work better than a single ambiguous commit. In our deterministic fallback, we also explicitly filter out bare keywords like wip, test, or commit so low-signal commits don't get inflated.
The "AI Proposes, Human Approves" Philosophy: Ultimately, we intentionally designed SlothLog as a lazy-first accelerator rather than an auto-submitting black box. That’s why every timeslot card has an inline editable textarea and an expandable dropdown showing the exact underlying commits and short SHAs. The AI takes care of 90% of the cognitive overhead of drafting the sentence, but you can soften "Delivered X" to "In-progress investigation on X" in 2 seconds before clicking "Copy for Jira".
I took a peek at your evidence repo—the ground-truth binding you’re doing with Python AST / source references is really neat and adds that layer of determinism small LLMs often lack.
Best of luck with SlothPulse and Hacktoberfest Week 3! 🦥🚀
Thanks, Alessio! The session-grouping approach gives me a useful experiment to try on ambiguous commits. Small clarification: SlothPulse uses Python to validate source references and coverage, rather than AST analysis. As our example showed, a valid reference can still accompany an incorrect completion claim—so I agree that human review matters. Thanks for sharing your approach, and good luck! 🦥
Thanks for the clarification, Teng! That makes total sense—ground-truth referencing confirms where the code changed, but discerning semantic intent and completion state is exactly where human review still shines.
Really glad the session-clustering angle gives you an interesting experiment to explore. Looking forward to seeing where you take SlothPulse!
a local, offline PDF toolkit for large print files. It compresses, splits, merges and removes image backgrounds on your own computer. Background removal uses an open-source model, and Gemma through Ollama handles page labels.
The most useful finding wasn’t a speed number. It was a skip. A photo covering 91% of a page was ignored by my 90% rule, and the tool only said “1 not processed”. It didn’t say why, so I changed the message to name the reason and the option that fixes it.
I recently built Spotlight — a launcher for people who hate waiting!
It's designed to make accessing apps and tools faster and simplify everyday workflows. I'd love to explore how open-source AI and local models could make it even smarter and more privacy-friendly.
Loved this week's theme. Sloth energy is exactly what I was going for 🦥
I built Actum, an Android app where you say one messy sentence ("buy rice, wake me at 6 tomorrow, mark 4th Nov as my birthday") and Gemma 4, running on the phone, turns it into ready cards for your to-do list, alarm and calendar. One tap each and it's done. The app doesn't even have the internet permission, so nothing you share into it can leave your phone.
The part I'm proudest of: I wrote 100 fresh test sentences, froze them, and measured rules vs Gemma vs the hybrid on a real phone. I published the failures too (prompt injection and negations still slip through).
Engineering student exploring AI, web development, and open source. Learning by building projects, solving real-world problems, and collaborating with developers. Always curious, always growing.
Location
Belagavi, Karnataka, India
Education
B.Tech in Computer Science and Business Systems (CSBS), VTU-affiliated college, Belagavi
My Week 3 entry is SlothPulse: Git commits plus an optional traceback become a standup draft through local Gemma/Ollama, with source references for review.
The most useful finding wasn't a speed number: the model cited the right commit but changed ‘submission handoff’ into ‘submitted.’ I kept the raw output and correction in the repo—valid source IDs don't guarantee a correct interpretation.
Hi! I want to participate in the Week 3 “Slow and Steady” challenge, but my laptop doesn't have a GPU or enough storage to run models locally.
Would a project using an open-weight model through the Groq API qualify, even though inference runs on Groq's infrastructure rather than locally?
I plan to build a lightweight Python application and document the model, API, and architecture clearly. I want to make sure this approach satisfies the open-source AI requirement before submitting. Thank you!
what's the compute budget limit for each repo, and are there any constraints on using external gpu services. also, how will you evaluate model performance without inflating the scoring pipeline?
May the slow and steady win the race this time as well, like all the other times 🤣🥰
Best wishes to all those who'll participate 🤞
And just in case you need some more free credits: 👇
Grab $250 in Free Cloud VM Credits from Docker (Before Oct 31, 2026)!
Good luck!!!🥰
BTW, this challenge suits your DEV user name perfectly! May be you should join ... just saying. 🤣🥰
Maybe DEV got inspired by my Dev user name and came up with this challenge, lol! Just kidding 🤣By the way, DEV should launch a Halloween challenge too! 🎃👻
Thanks! 🥰
Just published my submission for Week 3: SlothLog (🦥)!
It’s an offline, privacy-first developer utility that inspects local git commits, clusters them into work sessions by ticket ID, and writes clean timesheet worklogs using local Gemma 2 2B (with Sentry tracing and a Render cloud demo). No more wasting 45 minutes every Friday afternoon trying to remember what you did on Tuesday!
Check it out here: SlothLog
Happy International Sloth Day and good luck everyone! 🦥✨
Hi Alessio! I read your write-up—the ticket grouping and editable hour controls are useful touches. I'm working on a nearby problem with SlothPulse: local Gemma drafts a standup from commits and an optional traceback, with Python binding each item to its source.
One failure surprised me: a summary cited the correct commit but turned ‘submission handoff’ into ‘submitted.’ Source validation passed, yet the meaning still needed correction. Have you seen Gemma 2 2B turn tentative work into a completed accomplishment, especially with short ‘wip’ commits? Curious how you handle that before copying a worklog.
I preserved the raw example here if useful: github.com/TengByte/slothpulse/tre...
Good luck with the challenge! 🦥
Hi @tengbyte! Thanks so much for reading and for the thoughtful question! SlothPulse sounds like a fantastic concept—binding standup bullets to source commits and tracebacks is super smart.
You hit on a very real challenge: compact models like Gemma 2 2B definitely suffer from an inherent "completion bias". Because summarization fine-tuning heavily associates commit logs with deliverables, the model loves to turn tentative phrasing like "wip: auth handoff" into triumphant assertions like "Completed auth implementation."
Here’s how we approach and mitigate this in SlothLog:
0.2to minimize creative extrapolation and keep Gemma tightly tethered to the actual commit verbs.wip refactoring,investigating edge case,drafting tests), the chronological timeline helps anchor the reality of the work better than a single ambiguous commit. In our deterministic fallback, we also explicitly filter out bare keywords likewip,test, orcommitso low-signal commits don't get inflated.I took a peek at your evidence repo—the ground-truth binding you’re doing with Python AST / source references is really neat and adds that layer of determinism small LLMs often lack.
Best of luck with SlothPulse and Hacktoberfest Week 3! 🦥🚀
Thanks, Alessio! The session-grouping approach gives me a useful experiment to try on ambiguous commits. Small clarification: SlothPulse uses Python to validate source references and coverage, rather than AST analysis. As our example showed, a valid reference can still accompany an incorrect completion claim—so I agree that human review matters. Thanks for sharing your approach, and good luck! 🦥
Thanks for the clarification, Teng! That makes total sense—ground-truth referencing confirms where the code changed, but discerning semantic intent and completion state is exactly where human review still shines.
Really glad the session-clustering angle gives you an interesting experiment to explore. Looking forward to seeing where you take SlothPulse!
Have a great Hacktoberfest, and happy hacking! 🦥✨
Well, I'm not gonna be left out even though this is my first submission and I had to rush things just to meet up.
dev.to/thesonofraj/geosim-lab-a-vi...
My Week 3 entry is
PDFme
a local, offline PDF toolkit for large print files. It compresses, splits, merges and removes image backgrounds on your own computer. Background removal uses an open-source model, and Gemma through Ollama handles page labels.
The most useful finding wasn’t a speed number. It was a skip. A photo covering 91% of a page was ignored by my 90% rule, and the tool only said “1 not processed”. It didn’t say why, so I changed the message to name the reason and the option that fixes it.
Write-up, code and demo: dev.to/parnita_singh_73778510977/p...
I’d welcome feedback from anyone who has removed a background from a photo that fills the whole page. 🐘
I almost didn't make it due to time crunch
but its done now dev.to/0shuvo0/im-too-lazy-to-writ...
I recently built Spotlight — a launcher for people who hate waiting!
It's designed to make accessing apps and tools faster and simplify everyday workflows. I'd love to explore how open-source AI and local models could make it even smarter and more privacy-friendly.
🔗 Check out my project: dev.to/chndrapratapsingh/spotlight...
I'd appreciate your feedback! 🦥💻
This is my 3rd hackathon and again submitted at the last minute: 🫣dev.to/mohammad515253/sloth-migrat...
Loved this week's theme. Sloth energy is exactly what I was going for 🦥
I built Actum, an Android app where you say one messy sentence ("buy rice, wake me at 6 tomorrow, mark 4th Nov as my birthday") and Gemma 4, running on the phone, turns it into ready cards for your to-do list, alarm and calendar. One tap each and it's done. The app doesn't even have the internet permission, so nothing you share into it can leave your phone.
The part I'm proudest of: I wrote 100 fresh test sentences, froze them, and measured rules vs Gemma vs the hybrid on a real phone. I published the failures too (prompt injection and negations still slip through).
Write-up: dev.to/aritra_maity_e3bce29864e8/a...
Code: github.com/AM2004legend/ActIum
Would genuinely love feedback, especially from anyone running models on-device. Good luck everyone!
Looking forward to seeing all the submissions.
Excited for the upcoming events!!!
Slow and steady wins always
OHHHHHH~~！
祝所有参赛者好运
My Week 3 entry is SlothPulse: Git commits plus an optional traceback become a standup draft through local Gemma/Ollama, with source references for review.
The most useful finding wasn't a speed number: the model cited the right commit but changed ‘submission handoff’ into ‘submitted.’ I kept the raw output and correction in the repo—valid source IDs don't guarantee a correct interpretation.
Write-up, code, and a 48-second narrated demo: dev.to/tengbyte/slothpulse-i-wrote...
I'd welcome feedback from anyone testing local models on ambiguous commit messages. 🦥
Hi! I want to participate in the Week 3 “Slow and Steady” challenge, but my laptop doesn't have a GPU or enough storage to run models locally.
Would a project using an open-weight model through the Groq API qualify, even though inference runs on Groq's infrastructure rather than locally?
I plan to build a lightweight Python application and document the model, API, and architecture clearly. I want to make sure this approach satisfies the open-source AI requirement before submitting. Thank you!
nice!
why can't i redeem elevenlab promo code
Good luck to us all, i will try my best and shall be a winnerrrrr
great
May be this would help others new to this to learn open source #ai #open source #hactoberfest
Hello everything this is Narender from Jaipur , India. Hope you all are okay!
good
what's the compute budget limit for each repo, and are there any constraints on using external gpu services. also, how will you evaluate model performance without inflating the scoring pipeline?
hii