Yitumulin's Space
WHOAMI
/
PROJECT
/
CONTACT
@yitumulin
PROJECTS &
RESEARCH
Structured in a product mindset.

Each project below is presented by Problem, Strategy, Outcome, and Impact so people can quickly read what I solved, how I executed, and what measurable value was delivered.
Filter by focus:
LLM & Product Engineering
LLM Agent
WeClone End-to-End Deployment
Problem A complete local agent deployment was required, but the path from raw data to stable service was fragmented.
Strategy Built a staged deployment path covering environment setup, data prep, fine-tuning config, API startup, and chatbot validation.
Outcome Shipped a working local service where model output could be consumed through a complete conversational interface.
Impact Turned an experimental stack into an executable product workflow with verifiable end-to-end behavior.
WATCH ON DOUYIN ->
AI Agent Product (2026)
Yitumulin Agent for Personalized Planning
Problem Generic chat agents were weak in long-term memory recall, user-specific consistency, and execution-ready planning.
Strategy Designed a full chain from intent parsing, slot clarification, and memory retrieval to multi-agent coordination, generation, and tool execution.
Outcome Built a reusable personalized agent prototype with semantic-entropy-based follow-up logic for fitness and companion-planning scenarios.
Impact In 100+ rounds of validation, memory hit rate improved about 35%-40%, key-slot completion improved about 40%, and actionable-plan quality improved about 30%.
Provincial Innovation Project
NiuGuanHui Smart Cattle Management System
Problem Small and mid-size cattle farms lacked stable digital workflows for identity tracking, feeding optimization, and data-driven shipment decisions.
Strategy Led product scoping and technical validation for a non-contact CV-driven workflow covering profile creation, weight estimation, risk alerts, and operation dashboards.
Outcome Delivered a deployable solution with two core algorithm modules and application endpoints connected to real farm management scenarios.
Impact Collaborated with 9 farming enterprises and supported deployment in 3 leading sites, while improving detection quality and reducing manual operation cost.
AI Creator Track
AI Learning Content Creation (Douyin 2.5k)
Problem Practical AI engineering knowledge is often too technical for general learners to apply in daily study or product practice.
Strategy Continuously transformed agent workflows, LoRA/SFT practice, and deployment knowledge into short-form explainers with structured narrative and examples.
Outcome Built an ongoing educational content stream focused on LLM tooling and productized AI workflows.
Impact Reached about 2.5k followers and strengthened repeatable communication ability between engineering details and product-level understanding.
LLM System
MiniMind Full-Pipeline Reproduction
Problem Needed an end-to-end LLM training pipeline that could reliably run on a single GPU and be resumed after interruptions.
Strategy Rebuilt pretraining, SFT, LoRA, and evaluation stages with checkpoint and config validation in a reproducible local workflow.
Outcome Delivered a full runnable pipeline from data to eval report with stable handoff steps and clear execution documentation.
Impact Lowered reproduction risk and iteration friction by making the training loop restart-safe and evaluation-ready on one machine.
LLM + Speech
WhisperMail
Problem Users needed a faster way to convert spoken intent into structured email drafts without heavy manual editing.
Strategy Connected speech recognition, prompt orchestration, and email protocol handling into one consistent interaction flow.
Outcome Delivered a usable prototype that transforms voice input into action-ready email outputs.
Impact Improved productivity path from "idea to send" and validated product feasibility for AI-assisted communication.
Product Build
Shapeville Geometry Education
Problem Traditional geometry exercises lacked progressive feedback loops and made student engagement difficult to sustain.
Strategy Designed a staged challenge architecture with scoring feedback, progression rules, and interactive desktop UI.
Outcome Built a complete Java/Swing learning product with guided tasks and measurable learning checkpoints.
Impact Converted abstract geometry practice into a structured product experience with higher task completion continuity.
Academic Research
Scientific Data (2025)
BECA Dataset
Problem Long-term cattle re-identification lacked a broad and standardized dataset suitable for robust benchmarking.
Strategy Co-designed data collection and evaluation protocol for multi-scene acquisition with reusable benchmark baselines.
Outcome Published BECA with 16,889 images covering 5,661 cattle and an aligned verification workflow.
Impact Enabled reproducible long-term identity research and provided a practical standard for downstream CV evaluation.
DOI: 10.1038/s41597-025-06326-5
Comp. & Electronics in Ag. (2025)
CRODNet
Problem Rotated object detection in complex agricultural scenes required lower compute cost without losing recognition quality.
Strategy Optimized network architecture and benchmarked against strong baselines for efficiency-quality trade-off validation.
Outcome Published a lightweight detection approach validated on real task settings.
Impact Achieved about 70% parameter reduction and 50% FLOPs reduction while preserving practical detection capability.
DOI: 10.1016/j.compag.2025.111160
Under Review (2026)
Narrative Complexity Analysis
Problem Narrative complexity in long-form Xianxia texts lacked quantitative indicators with interpretable semantics.
Strategy Designed semantic entropy metrics and statistical testing procedures to model narrative structure evolution.
Outcome Built a computational framework linking language features with measurable complexity signals.
Impact Provided a transferable NLP evaluation path for literary analytics and cross-text comparison.
Software Copyrights
Software Copyright (2025.06)
Contactless Cattle Growth Tracking System
Problem Cattle growth tracking relied on contact-heavy and fragmented processes, making long-cycle monitoring costly and unstable.
Strategy Designed a CV-driven non-contact pipeline for top-view detection and growth-state tracking with consistent data recording standards.
Outcome Completed the core algorithm module and delivered a registrable software system for practical farm scenarios.
Impact Strengthened contactless identification capability and improved reliability for downstream monitoring and management workflows.
Software Copyright (2025.10)
EEG Micro-Texture & Closed-Loop Hidden-Frequency System
Problem BCI scenarios required more robust signal feature extraction and closed-loop control to keep interaction stable in practical use.
Strategy Implemented micro-texture EEG feature extraction and integrated a closed-loop control chain for real-time signal-processing response.
Outcome Delivered a complete software module stack and finalized registration as an independent software copyright.
Impact Improved engineering readiness for brain-computer interaction pipelines and validated end-to-end control feasibility.