pro
Zacharie Petchokomani
Senior Java, AI & Cloud Systems Engineer
New to Contra
Zacharie is building their profile!
Vibe Coding Prompt & Rubric Design
Created prompt-writing and rubric-evaluation tasks for client-side web applications. The work involved designing clear, testable prompts that instruct AI models to build interactive, stateful front-end apps, then writing atomic pass/fail rubrics to evaluate the generated application from a screenshot and submitted HTML/JS code.
I worked on tasks across categories such as web applications, data visualization, simulations, games, timers, recipe managers, avatar generators, audio visualizers, and management dashboards. Each task required defining the application concept, specifying visible UI behavior, ensuring the app was self-contained and client-side only, and building fair evaluation rubrics with clear tags such as visual, layout, content, interaction, and state.
The work also required calibrating task difficulty so model performance would fall into the expected range, especially for medium and hard tasks. I refined prompts and rubrics to make requirements specific, measurable, and difficult enough to expose model weaknesses while remaining fair and evaluable.
0
17
VS Code AI Baselining Tooling Setup
Configured and used a VS Code-based AI evaluation workflow for coding and infrastructure benchmark tasks. The setup involved preparing local development tooling, installing Python/Docker/Git dependencies, configuring multiple AI coding models inside VS Code, and using standardized scripts to download tasks, run benchmark tests, capture trajectories, and record model performance.
The workflow supported both software engineering and infrastructure tasks, including SWE-Bench-style bug fixes/features in real codebases and Terminal-Bench-style Docker troubleshooting tasks. Each task required solving the same problem across multiple models, tracking AI-assisted changes through Git commits, running tests, reviewing model behavior, and documenting failure modes and performance differences.
The screenshot shows the project dashboard/tooling area where the required shared VS Code tooling package was provided for task execution.
0
23
RLHF Prompt & Response Annotation
Completed RLHF annotation work using SuperAnnotate, focused on evaluating and comparing two AI-generated responses to user prompts. The work required reviewing the full task context, including the final user prompt, conversation history, and both model responses, then assessing each response for factual correctness, reliability, completeness, reasoning quality, clarity, formatting, and overall usefulness.
For each task, I identified strengths and weaknesses for both responses, assigned quality scores, selected the stronger response, and wrote concise preference justifications aligned with RLHF evaluation standards. The work included technical and coding-related prompts across languages such as JavaScript, TypeScript, PowerShell, and C#, requiring both software engineering judgment and AI response evaluation expertise.
The analytics screenshot shows production activity for an RLHF Code Annotation project, including 48 total items, 46 completed items, 2 submitted to QC, and 38h 47m of annotation work.
1
37
French AI Task Flagging Quality Check
Performed French-language quality control for AI task-flagging workflows on SuperAnnotate. The work involved reviewing prompts, conversation history, model responses, and annotator decisions to determine whether tasks required special flags such as recent knowledge, expert legal/healthcare/STEM knowledge, unsafe content, external-source access, PII, language mismatch, or whether no flag applied.
As a QC reviewer, I evaluated annotator selections, approved or rejected their work, provided professional feedback, and wrote concise task-flagging rationales in English for French prompts. The project required careful judgment, guideline interpretation, multilingual comprehension, and consistency with the calibration framework.
The analytics screenshot shows completed QA production for the RLHF 2.4 Task Flagging project, including 36 total items, 28 completed items, 8 completed items passed audit, and 3h 27m of QA work.
1
37