GitTrace
A student productivity tool that transforms Git history into structured internship logbooks, originally built to reduce the repetitive documentation required during my POLINEMA internship.
System Overview
GitTrace started from a simple problem: I was tired of manually writing my internship logbook.
As a POLINEMA student, I needed to document my development activities during an internship, but most of the work was already reflected across Git commits, repository history, and day-to-day development progress. GitTrace was built to turn that fragmented history into a more structured and readable internship logbook, reducing repetitive documentation work while keeping the report grounded in actual development activity.
By extracting repository records directly and applying structured prompt boundaries, GitTrace formats verified activity trails into institutional Word documents (.docx), Excel spreadsheets (.xlsx), and print-ready PDFs—helping students stay compliant without fabricating unrecorded work.
Architectural & Technical Challenges
Translating raw commit streams into meaningful logbook entries presents real technical hurdles. Commits are often shorthand phrases rather than descriptive summaries, and daily engineering tasks can be scattered across multiple local repositories.
Furthermore, student developers need offline privacy so proprietary repository code never leaves their laptops. The system had to run smoothly on consumer hardware with local Ollama models, enforcing strict JSON schemas so AI acts strictly to make recorded history readable without hallucinating or inventing activities that did not occur.
Key Architectural Decisions
Direct Git CLI subprocess traversal
Executed native Git binary commands via Dart's Process.run rather than compiling heavy C-bindings (libgit2), achieving instantaneous multi-repo directory traversal across deep file trees.
Strict JSON infill and schema constraints
Enforced rigid prompt boundaries that force LLM responses into structured JSON arrays, allowing seamless programmatic fallback when local models experience token truncation.
Local encrypted draft persistence
Decoupled raw Git commit parsing from user-editable draft entries using local encrypted storage, enabling students to customize descriptions prior to document export.
Engineering Contributions
System Architecture & MVP
Designed the initial application architecture using Flutter and Riverpod state management, including local repository discovery and an interactive monthly calendar interface.
LLM Context & Prompt Engineering
Engineered the AI context preparation pipeline, structuring multi-repo commit batches into bounded JSON schemas to run on consumer GPU VRAM and keep summaries strictly anchored to actual commit data.
Draft Review & Manual Annotation
Implemented an editable draft timeline so students retain full manual control to review, refine, and add context to their activities before document export.
CI/CD & Multi-Platform Delivery
Configured automated GitHub Actions workflows compiling and releasing standalone native desktop installers for Windows and macOS.
Technologies & Core Stack
Flutter • Dart • Riverpod • Ollama • DeepSeek • Git CLI