Jang Jinseo · AI Native Product Engineer
I develop financial solutions and support their operation. For my team, I stored handover notes and Git snapshots of uncommitted changes so development could continue on another PC, and separated execution environments by task. I configured Tailscale access to comply with the policy requiring use of only the headquarters database and introduced this AI-native development environment to the team. For my own products, I own planning through implementation and deployment.
01 · Experience
2024.12 — Present
I develop Java logic and operations interfaces for CruzLink and support launches and operations for financial institutions. I designed, built and introduced a team development harness that connects development information with release records and automates recurring management tasks.
2024.01 — 2024.06
I developed a Java and Spring Boot backend for processing and storing IMU Bluetooth and Radar sensor data in a medical gait-analysis system, along with patient and examination APIs. I designed PostgreSQL schemas and used tests and refactoring to support maintenance. I also handled deployment to AWS EC2 and automated deployment with GitHub Actions.
Dankook University · Bachelor’s degree in Visual Contents, minor in Software · graduated August 2023
Naver Cloud AIaaS Developer Program: 1,004 hours (2023) · TABA AI & Big Data: 640 hours (2022)
COVID-19 high-risk prediction ML model project · TABA Final Project Excellence Award (2nd place) · 2022
Gyeonggi Youth Gap Year IT Outstanding Case · 2023
TOEIC 960 · September 7, 2024
KATUSA · Republic of Korea Army · Sergeant, completed service · January 2017 — October 2018
02 · Team workflow
I stored handover notes and Git snapshots of uncommitted changes so development could continue on another PC, and separated execution environments by task. I configured Tailscale access to comply with the policy requiring use of only the headquarters database. I built a development harness connecting team knowledge with recurring task automation and introduced it to a team of nine developers, with people checking screens and behavior after AI implementation and checks.
Policy required the headquarters database even for local tests. I used Tailscale routing to provide database access without connecting each PC directly to the headquarters LAN. Across four shared PCs, files, Tomcat settings and ports are separated by Git worktree while the database is shared. Git snapshots of uncommitted changes and handover notes support moving work between PCs.
I organized knowledge scattered across conversations and messages as one topic per note, with sources, review dates and conditions of use for people and AI to consult. I used GitLab CI and the harness to automate issue management and prepare patch-note drafts from issues, merge requests and code changes, then register tags and releases after human approval.
AI handles implementation, checks and screen evidence. The person doing the work checks screens and behavior against the issue’s intent before opening a merge request; a reviewer then examines the changes. I also designed and implemented records of possible causes, supporting evidence, uncertainty and disagreement to distinguish specification or design errors from Claude’s failure to implement agreed requirements.
In real work, a database lock caused an approval input to go unrecorded and the approval to be rejected. I documented the issue and specified preserving the original input time, accepting the approval and recording failures as completion criteria. A fix that stores inputs and failures outside the database before applying them was merged. The merge request reports 13 passing checks, including lock reproduction, and validation of the Windows PowerShell invocation.
Version bumps in every feature merge request repeatedly caused merge conflicts. I proposed recording feature changes separately and updating the version only in release merge requests. Update detection had already moved to commit hashes. The revised rules use one version source, regenerate related files and combine change notes at release time.
Evidence scope · Design, implementation, team rollout and approval-event and usage-collection records from real work are confirmed. Nine is the team size. Deployment to every PC, team-wide active-user counts, reduced recurrence and productivity impact remain unverified. As of October 21, 2026.
03 · Personal products
I own planning, interfaces, backend, AI integration and deployment for Chaewoon and ART WINDOW. I built both web products with AI and defined the criteria for their explanations and generated outputs.
Chaewoon
Chaewoon connects Saju readings with gemstone products. I set the goal of clearly explaining how Saju interprets a person’s temperament and current life phase, and tasked AI with developing sentence principles, tone for each screen and review criteria. The resulting voice guide calls for stating the conclusion before the reason and making the actor, action and object explicit.
I used an AI-written source map to connect brand criteria with the server’s source documents. I also adopted AI’s proposal to separate generation and review, so passing automated checks would not stand in for readability. Records cover separate AI review of 12 generated samples and checks of deployed files, responses, and mobile and desktop screens. Interpretation accuracy, recommendation quality, customer impact and productivity were not measured.
Visit ChaewoonRead the knowledge, delegation and review caseProduct priorities and scope
Sources, file ownership, deliverables
Relevant tests, code, screens, deployment
ART WINDOW
ART WINDOW recommends and generates art for a room photo. When a photographic example led to a graphic image, I found different medium values for the same style ID and aligned the generation catalog with the example’s medium. I later found another difference: existing examples used reference images, while actual generation used text.
I retained text input because I considered plagiarism risk difficult to rule out with reference images. I made the medium and style description shared inputs and required examples to be updated when the description changes. The AI-implemented A12 check reports mismatches between sentence hashes and example records. A person then regenerates the example, updates its record and reruns the check.
Evidence scope · The published comparison shows previous and current examples, not before-and-after results for the same request. Isolated-input checks confirm mismatch detection, rather than image or recommendation quality.
Visit ART WINDOWRead the diagnosis and validation implementationBoth products are built and deployed.
04 · Operations tool
I built an operations tool with AI that links products, sourcing prices, orders and development history to their original sources. When I select an existing sourcing unit price and enter an assumed amount in the same currency and unit, it recalculates offer prices and estimated costs for linked items with registered formulas and quantities.
Results can be traced to source records and the conditions used. Calculating with an assumed price is kept separate from an actual transaction.
Recalculating offer prices and estimated costs with an assumed unit price leaves actual prices and historical order amounts unchanged.
Evidence scope · Source linking and conditional impact queries are implemented. This does not establish reconciliation of actual sales, inventory or financial records.
05 · How I work
I give AI the source material, constraints and desired result. I document procedures that have been checked so AI can follow them in later tasks. I keep the basis for decisions accessible after the work is finished.
In a technical documentation design assignment, the proposal used the existing internal wiki as an entry point to policy documents, approved API specifications and original experiment records in Notion. AI handled research into conventions, detailed writing, designs, diagrams and checks.
Personal harness templateAn example built separately from the company implementation
I build AI services and team tools, drawing on my development and operations experience.
