Abhijith Mullachery

AI/ML • Middle • Chandigarh, India • UTC+05:30 • Fully remote • Relocation • Work permits: India • Available in 2 weeks

Software Developer

Python

Node JS

JavaScript

Model Architecture & Training: 4/10

Active 1 day ago

Invite to interview

Download CVCV

Overview

Technical skills

Projects

Timeline

Roles

Overview

A Python backend engineer (middle level) with practical experience building REST and real-time APIs and integrating LLM-based RAG pipelines. The strongest proven skill is building API-backed features with real-time updates, evidenced by the Django Channels consumer and the JavaScript WebSocket manager (TODO_App/todo/consumers.py and todo-frontend/src/services/websocketService.js). The public code shows limited production hardening around secrets, backoff strategies, structured observability, and formal API versioning or pagination schemes.

Phone

Technical skills

Languages

Python

Node JS

JavaScript

Python

Django

Requests

FastAPI

Django REST Framework

SQLAlchemy

Pydantic

Asyncio

Beautiful Soup

Uvicorn

AI/ML

LangChain

LangGraph

NumPy

Pandas

Scikit-learn

huggingface_hub

TF-Keras

BERT

Streamlit

NLTK

DevOps

Docker

Rest API

AWS

GitHub

Git

Other

PostgreSQL

Milvus

React.js

Matplotlib

Dotenv

Groq

Jupyter Notebook

Machine Learning

RAG

Classic ML

Embeddings

Prompt Engineering

Recommender Systems

Semantic Search

NLP

Semantic Search

Projects

CineMate

Created a content-based movie recommendation system for Indian films using cosine similarity over multiple movie attributes. Engineered the recommendation logic with scikit-learn and delivered a Streamlit web app for interactive, real-time suggestions. Included movie insight display to help users understand why recommendations match their preferences.

Dietrix.AI

Built an AI diet recommendation system with FastAPI using a RAG pipeline for personalized nutrition plans. Implemented retrieval and response generation with LangChain, vector embeddings, and LLM integration. Automated data ingestion via web scraping to keep recommendations up to date, and provided APIs for a React frontend to deliver real-time personalized results.

Timeline

Web Developer • Middle

Snakescript solutiions • Full-Time

May 2025 to Present 1 Year 5 Months

Python

Django

RAG

Software Developer • Middle

SnakeScript Solutions LLP • Full-Time

May 2025 to Present 1 Year 5 Months Mohali In office

Developed backend features for scalable web applications using Django REST APIs, authentication, and PostgreSQL with Django ORM. Integrated AI chat assistants into Django apps using LangChain and LangGraph with large language models. Built an AI-powered chatbot by implementing document ingestion and semantic retrieval workflows, using Milvus for knowledge-base search and Docker for containerized deployment.

Python

Django

Rest API

PostgreSQL

LangChain

LangGraph

Milvus

Docker

Chandigarh University (CU)

Master's Degree • Data Science

2023–2025 Chandigarh, India

Mangalore University

Bachelor's Degree • Mathematics

2020–2023 Mangaluru, Karnataka

Middle AI/ML Engineer Confidence: High ML Engineer

NLP-focused ML engineer at a middle experience level with a practical strength in building end-to-end sentiment analysis pipelines and fine-tuning pretrained language models. The strongest proven skill is an end-to-end Malayalam sentiment pipeline including TF-IDF and classical model experiments in Sentiment-analysis/ML.ipynb plus a BERT fine-tuning and serving pipeline shown in Sentiment-analysis/BERT_.ipynb and Sentiment-analysis/app.py. There is limited evidence of production-grade MLOps, experiment tracking, robustness testing, distributed training or novel research contributions.

Model Architecture & Training

4/10

How well models are designed and trained

Practical model training and fine-tuning skills demonstrated with a full BERT fine-tuning loop and classical ML training pipelines.

Evidence

Sentiment-analysis/BERT_.ipynb: training loop using BertForSequenceClassification, AdamW optimizer and get_linear_schedule_with_warmup

Sentiment-analysis/ML.ipynb: end-to-end training of multiple classifiers and GridSearchCV for hyperparameter tuning

Data Pipeline & Feature Engineering

3/10

How data is prepared for models

Basic but correct data preprocessing and feature engineering for NLP including TF-IDF with a Malayalam stopword list and tokenizer-based batching for BERT.

Evidence

Sentiment-analysis/ML.ipynb: TfidfVectorizer setup with custom Malayalam stopwords and TF-IDF transform/save

Sentiment-analysis/BERT_.ipynb: preprocess_data using BertTokenizer and TensorDataset/DataLoader creation

Experimentation & Evaluation

3/10

How results are measured and tested

Standard experimentation and evaluation workflows with train/test splits, cross validation, GridSearchCV, metrics and visualization but no experiment tracking or reproducible run management.

Evidence

Sentiment-analysis/ML.ipynb: GridSearchCV and cross-validation for multiple models and printed classification reports

DL_lab/Confusion Matrix/Confusion Matrix&RoC curve.ipynb: ROC curve and confusion matrix plotting

Sentiment-analysis/BERT_.ipynb: validation loop computing loss and accuracy and printing classification_report

MLOps & Deployment

3/10

How models are shipped to production

Basic deployment steps shown - model saving and a small Flask prediction app - but no CI/CD, versioning, monitoring or scalable serving.

Evidence

Sentiment-analysis/app.py: Flask app that loads joblib models and serves predictions with input language checking

Sentiment-analysis/BERT_.ipynb: model.save_pretrained and tokenizer.save_pretrained usage

Computational Efficiency

2/10

How efficiently computing resources are used

Minimal efficiency work; uses batching and device selection but no profiling, quantization, distillation or GPU-memory optimizations.

Evidence

Sentiment-analysis/BERT_.ipynb: DataLoader with batch_size=32 and device selection (cuda if available)

Sentiment-analysis/BERT_.ipynb: use of scheduler to control training steps

Research Depth & Innovation

1/10

Depth of research and new ideas

No original research or novel architectures; implementations are correct but follow standard recipes without methodological innovations.

Evidence

Sentiment-analysis/BERT_.ipynb: standard fine-tuning of BertForSequenceClassification without custom layers

DL_lab/* notebooks: classical algorithm demos (PCA, Decision Tree, KMeans) showing instructional implementations rather than research contributions

Expertise

LLM• Middle

MLOps & Model Lifecycle• Middle

Industries

Media & Entertainment• Middle

Technologies

PostgreSQL• since 2025

LangGraph• since 2025

LangChain• since 2025

Milvus• since 2025

Groq

Embeddings

Scikit-learn• since 2026

Prompt Engineering

NLP• since 2026

huggingface_hub

Pandas

NumPy

Git

AWS

Docker• since 2025

RAG

BERT

NLTK

TF-Keras

Streamlit

Semantic Search

GitHub• since 2026

Semantic Search

Recommender Systems

GitHub• mentioned only

Sentiment Analysis• mentioned only

Recommendations

Add experiment tracking and reproducibility (WandB or MLflow) and log hyperparameter runs and artifacts from the BERT training loop.
Containerize and CI/CD the Flask service and the model artifact pipeline, and add model versioning and a simple inference test harness.
Introduce lightweight inference optimizations such as batching, TorchScript/ONNX export or quantization and measure latency and memory to meet production constraints.
Add unit tests, type hints, and configuration management for dataset paths and hyperparameters to improve maintainability and reproducibility.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

Abhijith7711 DL_lab Sentiment-analysis

Middle Backend Developer Confidence: Medium API Engineer

API Design

3/10

How well APIs are designed

Clear REST and FastAPI endpoints with JWT usage and basic error handling are present, but there is no evidence of versioning, idempotency keys, consistent pagination, or a formal error contract.

Evidence

TODO_App/todo/views.py: TodoViewSet and register_user use DRF viewsets and token issuance

TODO_App/todo_project/urls.py: JWT TokenObtainPairView/TokenRefreshView routing

Dietrix.AI/app/routes.py: FastAPI endpoint get_recommendations that invokes RAG chain

Data Layer & Database

3/10

Working with databases

There is a migration chain and database configuration with conn_max_age and health checks, showing schema evolution and some DB operational awareness, but no evidence of hand-tuned SQL, advanced transaction/isolation handling or N+1 elimination patterns.

Evidence

TODO_App/todo/migrations/0001_initial.py: initial schema migration

TODO_App/todo/migrations/0002_todo_description.py: follow-up migration adding a description field

TODO_App/todo_project/settings.py: DATABASES configured via dj_database_url with conn_max_age and conn_health_checks

Scalability & Performance

3/10

Handling load and speed

Scalability primitives are used - Redis channel layer for WebSocket, a persistent Chroma vector store and simple client-side reconnect throttling - but there is no documented cache invalidation, no queue-based decoupling for heavy ingestion, and limited evidence of measured load testing or sophisticated backoff/jitter strategies.

Evidence

TODO_App/todo_project/settings.py: CHANNEL_LAYERS configured to use channels_redis

TODO_App/todo-frontend/src/services/websocketService.js: reconnect logic, throttling, max attempts but no jitter/backoff policy

Dietrix.AI/app/rag_pipeline.py: Chroma vectorstore persistence and retriever config (k=10)

System Architecture

3/10

Overall system structure

Multiple coherent modules exist (Django REST + Channels, FastAPI RAG service, React frontends) indicating modular decomposition, but there is limited evidence of explicit inter-service contracts, reasoning about service boundaries, graceful degradation strategies, or configuration/secret management beyond.env checks.

Evidence

Dietrix.AI/app/main.py: FastAPI app and router import pattern

TODO_App/todo_project/asgi.py: ASGI application setup and channels routing

Dietrix.AI/app/routes.py: lazy-loading RAG chain and request parsing for the FastAPI service

Security & Auth

3/10

Protecting data and access

Authentication and input hygiene show awareness - JWT is validated in WebSocket consumers and password validators are enabled - but insecure defaults (CORS_ALLOW_ALL_ORIGINS True, SECRET_KEY fallback) and prompt-injection embedded in RAG prompts are present and reduce the security posture.

Evidence

TODO_App/todo/consumers.py: get_user_from_token uses rest_framework_simplejwt.AccessToken for WebSocket auth

TODO_App/todo_project/settings.py: SIMPLE_JWT and password validators configured, CORS_ALLOW_ALL_ORIGINS set to True

Dietrix.AI/app/routes.py: explicit LLM prompt construction embedded inside server code

Reliability & Observability

3/10

Stability and monitoring

There is basic observability and defensive coding - logging in consumers, Django LOGGING config, and try/except blocks across modules - but no structured correlation ids, metrics, traces, or robust retry/circuit-breaker patterns with reasoned timeouts are evident.

Evidence

TODO_App/todo/consumers.py: logger usage and try/except blocks around send/connect operations

TODO_App/todo_project/settings.py: LOGGING handler configuration to file

TODO_App/todo-frontend/src/services/websocketService.js: listener cleanup and reconnect scheduling

Expertise

Backend AI & LLM• Middle

Databases & Vector Storage• Middle

Messaging & Real-time• Middle

Microservices & API Architecture• Middle

Python• Middle

Industries

Artificial Intelligence• Middle

Health Care• Middle

Technologies

Python• since 2024 • Middle

Node JS• Middle

Rest API• since 2025

SQLAlchemy

FastAPI

Beautiful Soup

Django• since 2025

Asyncio

Pydantic

Uvicorn

Django REST Framework

Requests

Dotenv

Stack• mentioned only

Recommendations

Build and maintain REST and real-time API services that require JWT-based auth and WebSocket-driven updates, especially prototypes and MVPs.
Implement and extend RAG/LLM pipelines and ingestion workflows for domain-specific assistants, including vector store persistence and retrieval tuning.
Work on data ingestion and web-scraping modules that feed downstream ML systems, focusing on extraction, cleaning and document splitting.
Harden services for production: add secure secret management, structured tracing/metrics, robust retry/backoff with jitter, and formal API versioning and pagination.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

Abhijith7711 TODO_App Dietrix.AI

Junior Data Scientist Confidence: Medium ML Practitioner

A junior ML practitioner focused on hands-on implementation of classical machine learning algorithms and educational experiments in Python. The strongest proven skill is building and visualizing basic ML workflows, evidenced by the Decision Tree notebook (DL_lab/Car_purchase(DT_Clf)/Car_purchase(DT_Clf).ipynb) which trains and plots a DecisionTreeClassifier and reports accuracy. There is little evidence of production engineering, deployed model pipelines, data engineering at scale, or rigorous statistical analysis in the public artifacts.

Statistical Rigor

2/10

Correct use of statistics

Basic use of classification metrics (confusion matrix, ROC/AUC) and an EM implementation are present, but there are no assumption checks, uncertainty quantification, hypothesis testing with proper controls, or discussion of statistical limitations.

Evidence

DL_lab/Confusion Matrix/Confusion Matrix&RoC curve.ipynb: ROC curve, AUC and ConfusionMatrixDisplay usage

DL_lab/PCA/PCA.ipynb: StandardScaler use and PCA visualization (no statistical assumption discussion)

Data Wrangling & Cleaning

2/10

Preparing and cleaning data

Minimal data wrangling is present (simple synthetic datasets, basic mapping and scaling). There is little to no evidence of robust cleaning, missing-value strategy, provenance tracking, or leakage prevention.

Evidence

DL_lab/Car_purchase(DT_Clf)/Car_purchase(DT_Clf).ipynb: mapping categorical 'Income' to numeric via df['Income'].map

DL_lab/PCA/PCA.ipynb: StandardScaler usage for preprocessing before PCA

Exploratory Analysis & Visualization

3/10

Exploring and visualizing data

Visualizations are used to illustrate results (PCA scatter, decision tree plot, ROC curve, confusion matrix) but most plots are presented without in-depth written interpretation or storytelling linking findings to concrete questions.

Evidence

DL_lab/PCA/PCA.ipynb: scatter plot of PC1 vs PC2 colored by target

DL_lab/Confusion Matrix/Confusion Matrix&RoC curve.ipynb: confusion matrix plot and ROC curve

DL_lab/Car_purchase(DT_Clf)/Car_purchase(DT_Clf).ipynb: plot_tree visualization of trained DecisionTreeClassifier

Predictive Modeling

3/10

Building models that predict

Several classical models and a from-scratch PCA/algorithm experimentation style are implemented (Decision Tree, Logistic Regression, PCA). However, evaluation practice is limited to single train/test splits and toy/synthetic data with no robust cross-validation, calibration analysis, error breakdowns or production-readiness.

Evidence

DL_lab/Car_purchase(DT_Clf)/Car_purchase(DT_Clf).ipynb: DecisionTreeClassifier training and prediction with train_test_split

DL_lab/Confusion Matrix/Confusion Matrix&RoC curve.ipynb: LogisticRegression model trained and evaluated with ROC/AUC

Business Insight & Impact

1/10

Turning analysis into business value

There is little evidence of business framing, cost-sensitive reasoning, or actionable recommendations; notebooks are educational and algorithm-focused rather than tied to concrete business metrics or impact.

Evidence

DL_lab/Car_purchase(DT_Clf)/Car_purchase(DT_Clf).ipynb: simple objective statement creating a synthetic purchase dataset (educational framing)

Reproducibility & Notebook Hygiene

2/10

Clean, repeatable analysis

Notebooks include some reproducibility hints (random_state, seeds) and metadata, but there is no pinned environment, dependency file, data versioning, pipeline/code modularization, or CI/test artifacts.

Evidence

DL_lab/Confusion Matrix/Confusion Matrix&RoC curve.ipynb: np.random.seed(0) and train_test_split(random_state=42)

DL_lab/Car_purchase(DT_Clf)/Car_purchase(DT_Clf).ipynb: train_test_split(random_state=42) and model.fit usage

Technologies

Classic ML

Jupyter Notebook

Matplotlib

Machine Learning• since 2026

Recommendations

Develop educational prototypes and algorithm implementations that teach classic ML concepts, especially small-scale from-scratch implementations in NumPy and scikit-learn.
Work on prototyping and evaluation of supervised models for small to medium datasets, focusing on improving validation (cross-validation, calibration) and error analysis.
Contribute to or build reproducible notebooks and lightweight pipelines (requirements.txt, notebook-to-module refactor, simple CI) to bridge from experimentation to production readiness.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

Abhijith7711 DL_lab Coursera