Applied Machine Learning Engineer @ Mavenir
Susanta Baidya
I build production-grade LLM, retrieval and agentic systems — and care about why they work.
01About
Applied Machine Learning Engineer on Mavenir's Data & AI R&D team, with 1+ year of full-time industry experience building production-grade intelligent systems for telecom: document-to-knowledge pipelines, retrieval and agentic services, and fine-tuned domain models running on in-house GPU infrastructure.
My work sits at the intersection of LLMs, agentic AI, fine-tuning, retrieval and NLP, with a strong pull toward optimization, statistical learning and the mathematical foundations underneath it all.
I care about first-principles reasoning, understanding why a model behaves the way it does, and shipping systems that hold up outside a notebook.
- Focus
- LLMs · Agentic AI · Retrieval · Fine-tuning · NLP
- Based in
- Bengaluru, India
- Education
- MSc, Artificial Intelligence & Machine LearningIIIT Lucknow · 2023 — 2025
- BSc (Honours), StatisticsMaharaja Bir Bikram College · 2020 — 2023
02Experience
Mavenir
Apr 2026 — Present
Member of Technical Staff 1 · Data & AI R&D
Developing a Knowledge Base (KB) as a product.
Solulever
Jul 2025 — Apr 2026
Associate Software Engineer
Contributed to Brabo AIQ, Solulever's industrial IoT analytics platform.
- Developed AI-driven boiler efficiency optimization and predictive analytics.
- Built agentic LLM workflows with LangGraph and vector search.
- Delivered real-time operational recommendations through FastAPI services.
- Python
- FastAPI
- LangGraph
- Vector search
- ML
DeepMatrix
Apr 2025 — Jul 2025
AI/ML Intern
Developed and optimized YOLO-based computer vision pipelines for industrial defect detection.
- Improved model accuracy and inference speed.
- Reduced false positives in defect detection.
- YOLO
- Computer Vision
- PyTorch
Climate Resilience Observatory (CRO), Govt. of Uttar Pradesh
Sep 2024 — Apr 2025
Data Science Intern
Developed an end-to-end thunderstorm prediction system for Uttar Pradesh.
- Combined multi-source meteorological data into a single forecasting pipeline.
- Trained Bi-LSTM models that generate 12-hour forecasts.
- Supported early warning and disaster preparedness.
- Bi-LSTM
- Time series
- Python
03Projects
04Skills
LLMs & agents
- LLMs
- RAG
- Agentic AI
- LangChain
- LangGraph
- MCP
- Hugging Face
- Fine-tuning (LoRA)
Machine learning
- Python
- PyTorch
- TensorFlow
- Scikit-learn
- Computer Vision
- Time Series Forecasting
- NLP
Serving & infrastructure
- FastAPI
- Flask
- Docker
- Vector Databases
- PostgreSQL
- Git