0%
ML Engineer · MLOps

Aditya Srivastava

Based In

New Delhi, India

Status

Open to Opportunities

PyTorch llama.cpp Contrastive Learning QLoRA AWS Docker FastAPI NVIDIA Jetson MLflow Edge AI OpenCV Python
PyTorch llama.cpp Contrastive Learning QLoRA AWS Docker FastAPI NVIDIA Jetson MLflow Edge AI OpenCV Python
01 —Technical Skills
Core competencies
Self-Supervised & Contrastive Learning
Representation learning with custom contrastive loss functions, temporal pre-training, embedding space visualisation across epochs
Deep Learning Architectures
CNNs, BiLSTMs, Transformer fine-tuning, QLoRA domain adaptation (r=8, α=16), transfer & few-shot learning
MLOps & Pipelines
End-to-end ML pipeline design, MLflow / W&B experiment tracking, model versioning, CI/CD, GitHub Actions, Docker
Edge AI & Inference
NVIDIA Jetson AGX Orin & Orin NX deployment, llama.cpp CUDA acceleration, TTFT / tok/s benchmarking, resource-constrained systems
Cloud & Infrastructure
AWS (EC2, S3, IAM, VPC), Render, Vercel, Firebase (Auth, Firestore, Storage), Linux server management, secrets & deployment automation
Data Science & Visualisation
Feature engineering, time-series & irregular temporal data, synthetic data generation, Power BI, Pandas, NumPy, Scikit-learn
Programming & Frameworks
Python, C/C++, JavaScript, TypeScript, SQL, Bash — PyTorch, TensorFlow/Keras, OpenCV, FastAPI, Node.js, REST, OAuth2, Git
Architecture Designing
System architecture design, distributed inference pipelines, edge deployment strategies, scalable ML infrastructure planning
02 —Work History
2 positions
May 2026 – Present Current Role
HiSLM — Hierarchical SLM Inference Network
IIT (ISM) Dhanbad · Jharkhand, India
  • Developing a distributed inference architecture that routes queries across heterogeneous NVIDIA Jetson devices based on model capability tiers and task complexity for efficient edge deployment.
  • Implementing 4-bit and 8-bit quantized Small Language Models (SLMs) to enable on-device inference within the tight memory and power constraints of IoT edge hardware.
  • Building and evaluating domain-specific LoRA adapters for mining, medical, and agricultural knowledge domains — work conducted under Prof. Dr. Amogth Tarachand.
  • Designing a FastAPI-based client–server inference pipeline over LAN for low-latency on-device processing across Jetson AGX Orin and Orin NX nodes.
  • Specializing model capabilities across heterogeneous hardware tiers without full fine-tuning through domain-adapted LoRA modules tailored to each vertical.
May 2025 – Jun 2026 Completed
Freelancer
Outlier AI · Remote
  • Evaluating and refining LLM outputs — instruction-following quality, factual accuracy, and reasoning chain correctness across multiple project tracks.
  • Designing adversarial and edge-case prompts to stress-test model behaviour; contributing structured feedback to RLHF pipelines that reduced hallucination rates and improved calibration.
  • Producing high-quality labelled datasets through structured annotation and comparative ranking of model outputs, directly analogous to few-shot label curation in ML research.
  • Maintaining consistent inter-annotator agreement standards across sprints within a globally distributed async team.
03 —Projects
5 builds
01
PROJ / 01 · June 2026 - Present
RGVE — Response Generation Variability Engine
Python · TypeScript · FastAPI · Next.js · llama.cpp
Explores the LLM output possibility space by sampling across parameter configurations (temperature, top_p, persona, domain) and semantically clustering results. FastAPI backend with Metal-accelerated TinyLlama inference, priority-queue tree search for diverse response paths, and a Next.js frontend for real-time exploration of generation variability.
GitHub →
02
PROJ / 02 · Aug – Sep 2025
Lexical-Semantic Embedding Model
PyTorch · BiLSTM · Contrastive Learning · NLP
Deep learning embedding model combining lexical similarity and contextual semantic representation via a BiLSTM encoder. Custom contrastive loss pulls semantically similar pairs together and pushes dissimilar ones apart. Reproducible training pipeline with real-time embedding space visualisation across epochs.
GitHub →
03
PROJ / 03 · Sept 2025
Integriti — AI-Driven Security Defense Platform
Python · Transformers · Graph Neural Networks · Docker · Federated Learning
Multi-layered framework detecting, analyzing, and mitigating LLM misuse in hostile information operations. Real-time transformer-based classifiers (>90% accuracy) identify AI-generated content across text and multimedia. Forensic watermarking and fingerprinting trace outputs to source models and operators. Graph-based threat intelligence maps disinformation clusters and propagation chains. Privacy-preserving architecture via federated learning and differential techniques for national security defense.
GitHub →
04
PROJ / 04 · Dec 2025 – Feb 2026
Cloud-Deployed AI Automation Agent
AWS EC2 · n8n · Google APIs · Firestore · Linux
Event-driven AI automation system on AWS EC2 using n8n orchestration, integrating Telegram, Gmail, and Google Gemini APIs with Firestore for persistent state. Intent classification, multi-branch routing, and automated email pipelines; full deployment lifecycle managed on Linux infrastructure.
05
PROJ / 05 · May 2026 – June 2026
Synthrun Mail — Secure Multi-Tenant Email Platform
Node.js · Firebase Auth · Firestore · Nodemailer · Render · Vercel
Production-grade multi-tenant email platform for synthrun.site — decoupled static frontend on Vercel and Node.js REST backend on Render. Firebase Auth with server-side ID token verification, per-user Firestore mailbox isolation, SMTP relay via Nodemailer + Brevo with SPF/DKIM/DMARC, and infrastructure-as-code deployment via render.yaml.
GitHub →
04 —Made for Fun
5 side projects
FUN / 01
Pathetic — Pseudo-Code Programming Language
Python · Shell · Perl
A beginner-friendly pseudo-code-based programming language with its own interpreter. Supports variables, arrays, control structures (if-then-else, while, for), f-strings, arithmetic/comparison/logical operators, and more. Distributed as a pip package and Arch Linux PKGBUILD with CI/CD via GitHub Actions.
GitHub →
FUN / 02
Lunaris OS — Device Tree for Pixel 7
Makefile · C++ · Python · Shell · Android
Open-source contribution to Lunaris OS — a custom Android ROM. Maintained the device tree for Pixel 7 (panther) and Pixel 7 Pro (cheetah) with 4,600+ commits covering audio, Bluetooth, NFC, WiFi, sepolicy, and power optimization for the Google Tensor G2 platform.
GitHub →
FUN / 03
Stepheny — D&D Discord Bot
Python · Discord.py · Docker
Feature-rich Discord bot for Dungeons & Dragons — dice rolling, combat initiative tracking, character and HP management, NPC/tavern/weather generation, and YouTube music playback in voice channels.
GitHub →
FUN / 04
CompilerDesign-PyLib — Compiler Design Library
Python · HTML
pip-installable Python library covering the entire Compiler Design syllabus — lexical analysis, symbol tables, LL(1)/LR(0) parsing, shift-reduce, three-address code, DAG construction, and more. Every function returns plain dicts/strings with matching pretty-printers.
GitHub →
FUN / 05
DockHub — Docs, Sheets & Slides
Flutter · Dart · Firebase · Google APIs
Neo-brutalist Flutter app that unifies Google Docs, Sheets, and Slides — authenticates with Google, browses Drive, and provides built-in rich text, spreadsheet, and slide editors with offline drafts and file receiving from other apps.
GitHub →
05 —Education
2 degrees
01
B. Tech Computer Science
SRM Institute of Science and Technology KTR
Expected May 2027 · Tamil Nadu, India
Machine Learning · Deep Learning · Computer Vision · Data Structures & Algorithms · Cloud Computing · Probability & Statistics
02
10+2 Higher Secondary
Dhanbad Public School
Completed 2022 · Dhanbad, Jharkhand
PCM-IP (Physics · Chemistry · Mathematics · Informatics Practices)
06 —Certifications
1 credential
CERT / 01
Oracle Cloud Infrastructure Foundations Associate
Oracle Corporation
ID: 325117116OCI25DCFA · Valid Dec 2025 – Dec 2027
+
More coming
+
In progress
07 —Achievements
2 milestones
Sep 2025 – Jan 2026
Google Gen AI Exchange Hackathon 2025
Google Cloud · Prototype Contributor
Built a working GenAI prototype for "Generative AI for Demystifying Legal Documents," demonstrating applied LLM integration and cloud-based solution development under competitive constraints.
Jul – Sep 2025
OpenAI × NxtWave Hackathon
State-Level Finalist · Nominated for National Round
Qualified at state level with a competitive AI solution, nominated for the national round — demonstrating applied problem-solving and rapid AI prototyping under deadline pressure.
08 —Languages
4 spoken
English
Professional Working Proficiency — fluent in technical, academic, and business contexts
Hindi
Native — full fluency across all registers, spoken and written
German
Elementare Kenntnisse — grundlegende Lese- und Konversationsfähigkeiten, lerne aktiv weiter
Russian
Элементарный уровень — базовое понимание и разговорные фразы

Let's build
something
intelligent.

Get In Touch

Location

New Delhi, 110091

Let's build
something
intelligent.

Send me a message and I'll get back to you.