For employers

Hire this AI Trainer

Sign in or create an account to invite AI Trainers to your job.

LoginCreate account
Invite to Job
P
Parthiv Prakash

Parthiv Prakash

AI Data Annotator – LLM & Code Systems

India flagVellore, India
$15.00/hrEntry LevelOther

Key Skills

Software

Other

Top Subject Matter

Artificial Intelligence – Prompt Annotation & LLM Evaluation
Software Development – Code Annotation & Function Calling
Finance – AI Agent Training & Trading Systems

Top Data Types

TextText

Top Task Types

Function CallingFunction Calling
Prompt + Response Writing (SFT)Prompt + Response Writing (SFT)
Evaluation/RatingEvaluation/Rating
Question AnsweringQuestion Answering
ClassificationClassification
Text SummarizationText Summarization
Computer Programming/CodingComputer Programming/Coding

Freelancer Overview

a big part of my ML projects has been getting the data right before any model sees it. On the NASA C-MAPSS project, I processed raw sensor streams from 100+ engine units, engineered rolling statistics and slope features by hand, and used PCA to compress the feature space by 70-80% — all without any labeled failure data. That kind of unsupervised pipeline requires you to really understand what signal looks like versus noise, which is fundamentally the same judgment call that good data labeling demands. On the AI side, I've built and prompted multi-agent LLM systems where the quality of outputs depended heavily on how well I structured inputs, defined tool interfaces, and caught bad model behavior early. Working with DeepSeek, Gemini, Grok, and Ollama across the same system taught me how differently models respond to the same prompt, and how much that matters when you're building something reliable. That experience of shaping data and model behavior from both ends is what sets me apart. Education includes Master of Technology, Vellore Institute of Technology (2028).

Entry LevelEnglishHindiMalayalamTamil

Labeling Experience

Multi-Agent LLM System – Prompt Engineering & Function Calling

TextPrompt Response Writing SFT
Built a multi-agent autonomous trading system with four LLM-driven agents, each with a distinct investment personality and decision pipeline. A core part of the work involved crafting and refining prompts for each agent so their outputs stayed consistent, on-strategy, and tool-call ready. Integrated MCP (Model Context Protocol) to expose live tool backends — accounts, market data, and push servers — meaning every agent response had to be structured well enough to trigger the right function calls reliably. Worked across five different LLM backends (DeepSeek, Gemini, Grok, Ollama, OpenRouter) and observed firsthand how the same prompt produces wildly different outputs across models, which shaped how I wrote and evaluated responses. The work sat right at the intersection of prompt engineering, structured output design, and model evaluation.

Built a multi-agent autonomous trading system with four LLM-driven agents, each with a distinct investment personality and decision pipeline. A core part of the work involved crafting and refining prompts for each agent so their outputs stayed consistent, on-strategy, and tool-call ready. Integrated MCP (Model Context Protocol) to expose live tool backends — accounts, market data, and push servers — meaning every agent response had to be structured well enough to trigger the right function calls reliably. Worked across five different LLM backends (DeepSeek, Gemini, Grok, Ollama, OpenRouter) and observed firsthand how the same prompt produces wildly different outputs across models, which shaped how I wrote and evaluated responses. The work sat right at the intersection of prompt engineering, structured output design, and model evaluation.

2026 - 2026

ML Feature Engineering & Unsupervised Data Pipeline – NASA C-MAPSS

TextClassification
Processed and engineered training-ready data from raw sensor streams across 100+ aircraft engine units using the NASA C-MAPSS dataset. Built degradation trajectories from 20+ sensor channels using rolling statistics and slope-based features, reducing noise while preserving fault signals. Applied PCA to compress the feature space by 70-80% and ran unsupervised clustering to surface distinct engine health states — all without relying on labeled failure data. The output was a clean, structured dataset that could directly feed downstream classification or anomaly detection models.

Processed and engineered training-ready data from raw sensor streams across 100+ aircraft engine units using the NASA C-MAPSS dataset. Built degradation trajectories from 20+ sensor channels using rolling statistics and slope-based features, reducing noise while preserving fault signals. Applied PCA to compress the feature space by 70-80% and ran unsupervised clustering to surface distinct engine health states — all without relying on labeled failure data. The output was a clean, structured dataset that could directly feed downstream classification or anomaly detection models.

2025 - 2025

Education

V

Vellore Institute of Technology

Master of Technology, Software Engineering

Master of Technology
2023 - 2028

Work History

T

Trainedbyparthiv

Founder and Fitness Coach

Vellore
2024 - Present