Manav Garg

Full-Stack Software Engineer · Machine Learning · Deep Learning · Data Science

About

I am a full-stack software engineer working in machine learning, deep learning, and data science. Over the past two years, across large companies and startups, I have cut pipeline runtimes by more than 80%, reduced LLM hallucinations by 70%, and lowered infrastructure costs by 22%. I build, deploy, and maintain systems end to end.

Interests

• LLM systems: RAG, LangChain/LangGraph pipelines, hallucination reduction
• Vision Transformers for medical and satellite imaging
• Multi-agent orchestration and self-hosted AI infrastructure
• MLOps: model deployment, async inference, AWS
• Getting ML into production, not just notebooks

Experience

• Backend AI Engineer, Vault22, Remote (Mar 2026 – Jun 2026)
  • Rebuilt the ETL pipelines and data lake from scratch on S3, with validation gates at every stage, raising data integrity 60%
  • Removed 95% of production query errors by redesigning data layout and indexes around real access patterns, and cut infrastructure cost 22%
  • Lifted model serving reliability 40% and cut compute overhead 30% by moving inference onto self-hosted Kubernetes
• Associate Data Scientist, IHX Pvt. Ltd. (A Perfios Company), Bangalore (Jan 2025 – Mar 2026)
  • Owned the tariff digitization initiative end to end, taking a five-day manual process down to a few hours
  • Reached 92% extraction accuracy on scanned and photographed documents by fine-tuning and benchmarking YOLO and Donut, running the annotation in Label Studio and the error analysis that selected the production model
  • Cut LLM hallucination over 70% with zero fine-tuning, evaluated against real production failures rather than a benchmark
  • Ran an async inference service at 10,000+ daily requests under 200ms

Internships

• AI/ML Intern, Intel Corporation, Chennai (May 2024 – Aug 2024)
  • Classified contract clauses and flagged anomalies at 94% accuracy with an end-to-end NLP pipeline on an event-driven architecture
  • Detected subtle fine-print deviations and fraud signals at 89% precision using semantic embeddings
• Full Stack & ML Intern, Tisac Pvt. Ltd., Remote (May 2024 – Jun 2024)
• Full Stack Developer Intern, Hogarth India (WPP), Gurugram (Aug 2023 – Oct 2023)
• Android & ML Intern, Wipro, Chennai (Aug 2023 – Nov 2023)

Selected Projects

• Orqestra. Self-hosted AI orchestration platform. Runs LLMs locally, sandboxes agent code, and streams live activity. Records the full trajectory of every run, each step, action, and result, and ships with evaluation tooling that replays runs across model and prompt versions, so a change can be shown to improve agent behaviour rather than just move the problem. Go, Python, Docker, Redis, and WebSockets, with AgentHive and LiteLLM behind an OpenAI-compatible API.
  Code
• HoneyML. LLM-driven adaptive SSH honeypot. Generates a convincing shell per attacker session and holds coherence across long adversarial interactions, mapping observed behaviour to MITRE ATT&CK so intrusion activity becomes structured data rather than raw logs. In research and build.
• LinkedIn Profile API. Structured profile data through LinkedIn's internal API, after establishing that the data is stripped server-side and cannot be scraped. Discovers the rotating endpoint version at runtime rather than hardcoding it, and returns explicit schema drift errors instead of plausible wrong data.
  Code
• Snippy. Short-form video generator with a web UI. Custom characters, background music, AI voices, animated subtitles, and auto-generated images.
  Code
• LuminSkin. Skin cancer classification with Vision Transformers. Benchmarked ViT and DeepViT in PyTorch, reaching 95%+ accuracy on lesion classification.
  Code
• Satellite Image Analysis (STIA-ViT). Land-use classification from satellite images. Benchmarked ViT against a Compact Convolutional Transformer, reaching 96% accuracy.
  Code
• KeyKeg. Password strength analyzer and manager. Scores strength, stores credentials, and generates passwords. The Random Forest strength model reached 93% accuracy.
  Code
• NSCHR. Health Management Information System web portal. Cut patient record retrieval time threefold; sold to clinics.
  Code
• Rekol. WhatsApp and web assistant for reminders, daily tasks, grocery lists, and notes.
  Code

Writing

I write about inference optimisation, agent architectures, evaluation, and LLM failure modes.
Posts on LinkedIn

Technical Skills

• Languages: Python, Go, C/C++, JavaScript/TypeScript, Java, SQL, Bash
• Machine Learning & AI: PyTorch, TensorFlow, Scikit-Learn, HuggingFace, Computer Vision, NLP, Vision Transformers (ViT), YOLO, OpenCV, Document AI (Donut)
• LLM & Generative AI: LangChain, LangGraph, LangSmith, RAG, Prompt Engineering, MCP, Agent Orchestration, Tool Calling, Evaluation Frameworks, Guardrails, LoRA/PEFT Fine-Tuning
• MLOps & Cloud: CI/CD, Model Deployment, Docker, Kubernetes, AWS (EC2, S3, SageMaker), ETL Pipelines, Jenkins, DeepSpeed ZeRO, Self-Hosted Model Serving, Monitoring and Alerting
• Data Engineering: ETL Pipeline Design, Data Lake Architecture, Spark, Hive, Kafka, Airflow, pandas, NumPy
• Backend & Web: FastAPI, Django, Node.js/Express, React/Next.js, REST APIs
• Databases & Tools: MongoDB, PostgreSQL, Redis, Firebase, Git, Linux, Label Studio

Education

B.Tech in Computer Science and Engineering, GPA 8.51/10
Vellore Institute of Technology, Chennai (Aug 2021 – Jun 2025)
Coursework: Data Structures & Algorithms, Machine Learning, Artificial Intelligence, Explainable AI, Linear Algebra, Multivariate Calculus, Operating Systems, Computer Networks, Probability & Statistics, OOP

Certifications

• Anthropic AI Safety Fellowship, reached the final interview round.
• Mathematics for Machine Learning: Linear Algebra, Imperial College London (May 2026). Credential ID 6SXRD8VFWQKG.
• Mathematics for Machine Learning: Multivariate Calculus, Imperial College London (Jul 2026). Credential ID KCRIC2735USG.