Streams of scanned documents and bounding boxes coalesce and resolve into a portrait of Febin Ahmed

DOCUMENT AI · COMPUTER VISION · PRODUCTION ML

I build machine-learning systems that read documents.

Research to production. Bangalore.

View case studies Email

// extracted from shipped systems

95%document classification — hybrid CV + NLP + layout90%key-value & table extraction — segmentation + rules−25%OCR errors, via custom training87%chart-generation accuracy — fine-tuned local LLM

01 · In action

In action

Real systems, running.

A digital menu-board wall concept rendering on a screen above a busy restaurant counter at dusk

SHIPPED · MULTI-TENANT SAAS

Restaurant DM

Design a menu wall in the browser; watch it change on the restaurant’s screens in real time.Live at Paragon Restaurants, Kerala.

  • stackFastAPI · Postgres · Redis · React
  • screensRaspberry Pi kiosks
  • updatesreal-time (SSE), offline-safe (IndexedDB)

Concept visual in the product’s design language.

A scanned page separates into glowing structured field and table layers

IN PRODUCTION · DOCUMENT AI

From scanned page to structured record

Hierarchical segmentation turns pages into key-value pairs and tables — at 90% accuracy in production.

Read the case study

02 · Selected work

Selected work

Four production systems from my R&D role at Ninestars — each one research-to-deployment, each with a number that held up in production.

Case study · Classification

Hybrid document classifier

One model wasn't enough: a classifier that fuses computer vision, NLP, and layout analysis to sort real-world documents at production scale.

accuracy
95%
signals
vision + text + layout
stage
production
  • PyTorch
  • OpenCV
  • Transformers
  • Docker
Read case study

Case study · Extraction

Key-value & table extraction pipeline

Hierarchical semantic segmentation plus a rule engine that turns scanned documents into structured key-value and tabular data.

accuracy
90%
manual work
−30%
stage
production
  • U-Net
  • Segmentation
  • Python
  • SQL
Read case study

Case study · OCR

Cutting OCR errors by a quarter

Custom training methods for OCR models on hard documents — a 25% error reduction that every downstream system inherited.

errors
−25%
method
custom training
stage
production
  • OCR
  • PyTorch
  • Data pipelines
Read case study

Case study · LLM systems

Multi-modal RAG chatbot on local GPUs

An open-source-LLM chatbot serving multi-document RAG, multi-agent responses, and chart generation — deployed entirely on local hardware.

graph-gen llm
87%
runs on
local GPUs
stage
deployed
  • Ollama
  • LLaMA-Factory
  • RAG
  • FAISS
Read case study

03 · Skills

Skills

Grouped by what I actually use them for.

Languages & frameworks

  • Python
  • Django
  • FastAPI

Web / product

  • React (SPA + PWA)
  • PostgreSQL
  • Redis
  • SSE
  • IndexedDB offline-first
  • Cloudflare Tunnel

ML / DL

  • PyTorch
  • TensorFlow
  • Keras

Computer vision

  • OpenCV
  • Faster R-CNN
  • U-Net
  • ViT
  • YOLO
  • LayoutLM

NLP

  • BERT
  • Transformers
  • LangChain

Generative AI / LLMs

  • LLaMA-Factory
  • Ollama
  • multi-modal chatbot design

Document AI / extraction

  • Hierarchical segmentation
  • OCR training
  • RAG

Data

  • Pandas
  • BeautifulSoup
  • SQL
  • MongoDB

Databases / search

  • FAISS
  • Elasticsearch

Visualization

  • Tableau
  • Plotly
  • Bokeh

Tooling

  • XGBoost
  • MLflow
  • Git
  • Docker
  • Pydantic
  • CrewAI

04 · Experience

Experience

Nov 2025 – Present

Senior Content Contributor, AI

Skillsoft · Bangalore

  • Building agentic workflows that auto-create teaching materials, slides, and voiceovers.
  • Pairing human expertise with AI systems to develop business solutions.
Sep 2024 – Oct 2025

Learning Specialist, AI/ML

Emeritus · Bangalore

  • Built AI-powered chatbots and automation pipelines to resolve learner course queries.
  • Built custom AI tools for automated grading, feedback generation, and student performance tracking.
  • Designed and delivered AI/ML curriculum — ML, deep learning, NLP, computer vision, LLMs — and launched a specialized LLM course covering fine-tuning, prompt engineering, and real-world applications.
Jul 2022 – Aug 2024

R&D Machine Learning Engineer

Ninestars Information Technologies · Bangalore

  • Led adoption of AI and computer vision for document analysis and extraction, owning projects from research to full-scale production.
  • Built a hybrid CV + NLP + layout document classifier reaching 95% accuracy, and a hierarchical-segmentation extraction pipeline reaching 90%.
  • Cut OCR model errors 25% with custom training; automated extraction workflows to reduce manual intervention 30%.
  • Designed and deployed a multi-modal RAG chatbot on local GPUs using open-source LLMs; led a team of data annotators, front-end devs, UI/UX designers, and DB managers.
→ see the four case studies
Apr 2022 – Jul 2022

Data Analysis Intern

Gameskraft · Bengaluru

  • Built and automated SQL queries powering customer-retention analysis and reporting for marketing.
  • Built predictive models in Python to forecast customer behavior, with dashboards in Tableau, Bokeh, and Plotly.

05 · About

About

I’ve spent the last ~3 years on one theme: making machines genuinely useful with messy, real-world information — most of it locked in documents.

At Ninestars I led R&D for document AI: classifiers that read pages the way people do (words, visuals, layout), segmentation pipelines that turn scans into structured data, OCR models trained until the error rate dropped by a quarter, and a multi-modal RAG chatbot that runs entirely on local GPUs. I owned these systems from research through full-scale production, and led the annotators, front-end developers, designers, and DB managers it took to ship them.

On the side I built and operate Restaurant DM — a multi-tenant digital menu-board platform running in production in Kerala: FastAPI backend, three React frontends, and Raspberry Pi kiosks that stay live through network drops.

Then I spent a year at Emeritus teaching what I’d been practicing — designing AI/ML curriculum across deep learning, NLP, computer vision, and LLMs. Today I’m at Skillsoft building agentic workflows that auto-create teaching materials — AI systems and education, converging.

M.Tech in AI & Data Science (Amrita Vishwa Vidyapeetham, 2022); B.Tech in Computer Engineering (MIT Pune, 2019). Based in Bangalore.Open to conversations about Document AI, computer vision, and production ML.

06 · Education & credentials

Education & credentials

Degrees

  • M.Tech, Artificial Intelligence & Data ScienceAmrita Vishwa Vidyapeetham, Coimbatore · 2022
  • B.Tech, Computer EngineeringMaharashtra Institute of Technology, Pune · 2019

Publication

  • Snake Intrusion Detection System using Optical Flowpublished Jun 2021

Certifications

  • LLM Mastery: Transformers and Generative AI — Sep 2024
  • LlamaIndex: Develop LLM Applications (Udemy) — Apr 2024
  • LangChain: Develop LLM-Powered Applications (Udemy) — Feb 2024
  • HackerRank SQL (Intermediate) — Mar 2022 · HackerRank Problem-Solving — Apr 2022
  • Python for Data Science (IBM) — Jul 2020

Earlier work

  • Link Prediction on the Facebook dataset (Kaggle)Neo4j, MLlib, PySpark · Jun 2020

07 · Contact

Contact

The fastest way to reach me is email. I’m in Bangalore (IST) and read everything.

Email me
Febin turns from his workshop desk and looks up