Open to AI Roles — Remote & India

Harshit Gavita

AI Systems Engineer · Agentic AI · Multi-Agent Orchestration
ML Infrastructure · LLM Integration · Low-Level Optimization

Building production-grade AI systems from the ground up — from autograd engines in JAX to 7-layer agentic safety architectures. 1st-year CSE (AI & DS) student with real shipped ML pipelines.

Building AI that
works in the real world

I'm a 1st-year B.Tech CSE (AI & DS) student at Nxtwave Institute of Technologies, Pune, deeply focused on AI systems engineering — not just using models, but understanding how they work at every layer.

From rebuilding Karpathy's micrograd from scratch in JAX to shipping production ML pipelines at CodeAlpha, I work at the intersection of low-level systems and applied LLM engineering.

My current obsession: agentic AI safety, multi-agent orchestration, and making AI systems reliable enough to actually deploy. I documented a real incident where a local AI agent crashed my workstation and proposed a 7-layer fix.

4
Production ML Pipelines shipped
>85%
Accuracy across all systems
10×
Cycle reduction in VLIW challenge
19
ML notebooks from scratch

Skills

End-to-end stack from low-level optimization to deployed AI products.

🐍
Programming Languages
PythonC++ JavaScriptNode.js HTML5/CSS3
🧠
ML / DL Frameworks
TensorFlowKeras JAXScikit-learn XGBoostCNNLSTM
🤖
AI & LLMs
Claude APIOpenAI API OllamaGemini API Prompt Eng.Agentic AI
⚙️
Deep Learning
Autograd from scratch Backpropagation Language Models Gradient Descent
📊
Data & NLP
PandasNumPy LibrosaMFCC OpenCVSentiment Analysis
🔧
Systems & Tools
VLIW / ILP Dockern8n FirebaseGit/GitHub Google Colab

Experience

Real production systems, shipped and evaluated.

Dec 2024
– Jan 2025
Machine Learning Training Intern
CodeAlpha · Remote
  • Designed and delivered 4 end-to-end ML pipelines — credit scoring, CNN handwritten character recognition, disease prediction, and speech emotion recognition — all achieving >85% accuracy within a 6-week timeline.
  • Engineered full ML lifecycle: data ingestion → preprocessing → MFCC feature extraction → model training → hyperparameter tuning → evaluation (Accuracy, ROC-AUC, F1-score).
  • Shipped 4 production Python modules using TensorFlow, Keras, Scikit-learn, and Pandas with clean API boundaries and reproducible inference pipelines.
2023
– Present
B.Tech CSE — AI & Data Science
Nxtwave Institute of Technologies, Pune
  • Specialization in Artificial Intelligence and Data Science.
  • Active in competitive programming, open-source contributions, and building personal AI projects alongside coursework.
  • Mumbai Hacks 2025 — Round 2 Qualifier (national-level hackathon).

Projects

From autograd engines to VLIW optimization — always close to the metal.

Systems

Anthropic Performance Takehome Challenge

Implemented KernelBuilder.build_kernel() for a simulated VLIW-style multi-core machine. Reduced execution cycles from 147,734 → ~14,425 — a ~10× improvement — while preserving exact output correctness.

VLIWILP PythonMemory Hoisting
~10×cycle reduction
ML

Credit Scoring Model

End-to-end credit risk pipeline with imputation, categorical encoding, and mutual-information feature selection. Benchmarked Logistic Regression, Random Forest, and XGBoost — achieved ROC-AUC ~0.91 on held-out test.

XGBoostScikit-learn PandasFeature Engineering
ROC-AUC 0.91on held-out test
DL

Speech Emotion Recognition

Extracted 40-dim MFCC features with Librosa. Trained LSTM achieving ~82% accuracy across 7 emotion classes with a fully reproducible TensorFlow/Keras end-to-end training and inference pipeline.

TensorFlowLSTM LibrosaMFCC
82%across 7 emotion classes

Published Work

Empirical research on autonomous AI agent safety — from a real incident.

Research Paper

Protection Toward Future (PTF): An Empirical Incident Study of Autonomous AI Agent Behaviour on a Personal Workstation

Documented a real incident where a locally deployed AI agent stack (Ollama + Kimi 2.5 + OpenClaw with a 9-layer PROMETHEUS autonomous system prompt) crashed a personal workstation. Proposed a 7-layer PTF safety architecture as a response.

Agentic AI Safety Architecture Ollama Multi-Agent Systems Empirical Study
7-Layer PTF Safety
1
Resource Monitoring
2
Action Sandboxing
3
Intent Verification
4
Loop Detection
5
Human-in-the-Loop Gates
6
Rollback Mechanisms
7
Emergency Kill-Switch
Ready to build
something serious?

Open to remote AI/ML engineering roles, research collaborations, and interesting problems in agentic AI.