Ankit Kumar Singhbuilding production ML,
one experiment at a time.
ML Software Engineer at Andor Communications. I fine-tune models, build RAG pipelines and AI agents, and ship them into real products used by people — not just notebooks.
Who I am
I'm a Machine Learning Software Engineer working on Andor Communications' AI product suite, where I fine-tune deep learning models for production use and build the infrastructure — RAG pipelines, LLM fine-tuning, AI agents — that makes them deployable, not just demoable.
My background is in Computer Science and Design from IIIT Delhi, where I spent four years moving between applied ML projects (classification, generative modelling, signal filtering) and the systems work needed to actually run them.
I care most about the gap between a model that works in a paper and one that works in production: calibration, evaluation, latency, and the unglamorous engineering that closes that gap.
Experience
Andor Communications Pvt Ltd
- Fine-tuned and implemented BiRefNet-based deep learning models for background image processing on custom datasets built in-house, optimized for production AI tools.
- Fine-tuned and optimized Hugging Face models on custom datasets; implemented LLM fine-tuning (LoRA / PEFT), RAG pipelines, argument-driven configurations, and AI agents for automation — with robust evaluation and scalable deployment.
Featured build
The latest thing off the workbench — pulled straight from GitHub.
DocuGuardian NEW
An AI document-protection agent that goes past summarizing. It reads loan agreements, insurance policies, and contracts across a whole workspace, surfacing hidden penalties, conflicting clauses, and upcoming deadlines before they turn into expensive mistakes.
- OCR → classification → structured extraction pipeline for PDFs and scanned documents
- Cross-document reasoning that catches contradictions and shared risk across uploads
- Dedicated risk-scoring and clause-severity models, grounded in the source clause
- Multilingual summaries and voice playback for non-English readers
Selected projects
Academic and independent work spanning classical ML, generative modelling, and applied systems.
Breast Cancer Classification
Compared five classical ML models for tumour classification, then used PCA to compress the feature space while preserving almost all of the signal.
Autoencoder-Driven Gen AI with Kalman Filter
A lightweight architecture pairing autoencoders with Kalman filters for text classification, sentiment analysis, and sequence modelling on memory- and compute-constrained systems.
Online Pharmacy
Rebuilt a CLI transaction platform end to end, automating invoicing and order workflows on top of a redesigned relational schema.
Stack
Languages
ML & Deep Learning
Backend & Infra
Databases
Dev Tools
Platforms
Core
Achievements & leadership
Get in touch
Open to ML engineering roles, research collaboration, and interesting problems. Type a message or use the mic to dictate it.
The fastest way to reach me is email — every message from this form opens directly in your mail app, addressed to me.