Sayantan Chaklader — AI & Data Science

EXPERIMENT LOG

Learned by building, not by theorem.

Sixteen months moving through statistics, classical ML, deep learning and NLP — every project shipped end to end and measured, not just modeled.

= starred on GitHub

GenAI · Automation

AI Job-Lead Pipeline

A daily n8n pipeline that collects new AI job listings open to India, filters them with cheap rule checks first, then has an LLM judge score each one into structured JSON and sends a shortlist to Telegram. Job text is treated as untrusted: a regex tripwire and a judge-side flag quarantine prompt-injection attempts, and nothing is applied for without me.

5 / 5Injection attempts caught · 98% India-eligibility accuracy on a 55-job eval

n8nGroqGoogle SheetsTelegram

View repo

GenAI · LLM

Research Agent

A LangGraph research agent that refuses to trust its own answers — evidence is extracted with exact-quote grounding, checked for sufficiency across bounded research iterations, and a citation only becomes "trusted" after passing both deterministic provenance validation and a separate semantic verification pass. Verification fails closed: no confident guesses, just a safe fallback with zero citations.

2,124 testsPassing · fail-closed grounding + semantic verification

LangGraphGeminiTavilySSRF-Safe Fetching

View repo

GenAI · Data Analysis

AI Data Analyst Agent

Upload a CSV, ask questions in plain English — the agent turns them into DuckDB SQL via Groq, then an independent SQLGlot-based validator checks every query is read-only before it ever touches the database. Follow-up questions now carry bounded session context, and DuckDB itself is sandboxed as a second layer of defense.

360 tests88.49% coverage · CI on every push · live on Streamlit Cloud

GroqDuckDBSQLGlotDocker

View repo

GenAI · LLM

RAG-Based Conversational AI

A retrieval-augmented generation pipeline that grounds every answer in a source document — chunked, embedded, stored in ChromaDB, and retrieved at query time — instead of letting the LLM improvise.

GroundedSays "I don't know" over fabricating an answer

LangChainChromaDBOpenAI EmbeddingsTop-k Retrieval

View repo

Deep Learning · NLP

Sentiment Analysis on Amazon Fine Food Reviews

LSTM network classifying reviews as positive or negative — built end to end from an 887K-review raw dataset down to a balanced, tokenized training set.

89.25%Test accuracy · F1 89.1%

LSTMTensorFlow/Keras620K+ samples10K-word vocab

View repo

Machine Learning · NLP

Netflix/TMDB Movie Recommendation System

Content-based recommender over 4,800+ movies — genres, keywords, cast, crew and overviews merged into per-movie "tags," then vectorized to surface the closest matches.

5,000-dimCosine similarity, top-5 recommendations

NLTKPorterStemmerCountVectorizerCosine Similarity

View repo

Machine Learning

CS:GO Match Outcome Prediction

Classification pipeline over competitive CS:GO match data — round dynamics, weapon economy, map and side — to forecast the winning team from mid-match state, not just pre-match odds.

Mid-matchPredicts from live game state, not pre-game odds

Logistic RegressionFeature EngineeringSeaborn EDA

View repo

Deep Learning

Handwritten Digit Recognition

CNN trained on MNIST, with data augmentation and L2 regularization to keep it from memorizing the training set.

99.7%Validation accuracy

CNNTensorFlow/KerasData Augmentation

Machine Learning

Census Income Classification

Ran a PyCaret model comparison to predict income bracket from census data — Gradient Boosting came out ahead.

85.8%Accuracy, vs. 73% baseline

PyCaretGradient Boosting

Machine Learning

Heart Disease Risk Prediction

Tuned decision tree on clinical data, with Seaborn heatmaps surfacing which indicators actually drove risk.

80.3%Accuracy

Decision TreeHyperparameter TuningSeaborn

Machine Learning

Flight Price Prediction

Regression pipeline forecasting ticket prices, refined through iterative feature engineering.

R² 0.985RMSE ≈ 2,779

RegressionFeature Engineering

Machine Learning

Bio-Signal Smoking Detection

Classification pipeline on biometric signal data to flag smoking status from physiological markers.

83.3%Best-model accuracy

ClassificationBiometric Data

Deep Learning · NLP

EN → JP Neural Machine Translation

Built a seq2seq LSTM encoder–decoder from scratch — not to ship a product, but to actually understand how translation models learn.

Seq2SeqLSTM encoder–decoder

LSTMEncoder-Decoder

Machine Learning

Image Classification Benchmark

GridSearchCV hyperparameter search across multiple Scikit-learn classifiers to compare approaches head to head.

GridSearchCVMulti-model search

Scikit-learnGridSearchCV

Data Analysis

HR Analytics — Hypothesis Testing

Normality testing and conditional filtering on employee satisfaction data to validate assumptions before modeling.

EDANormality + hypothesis testing

Hypothesis TestingStatistics

GENESIS STORY

From curiosity to code.

I grew up in Raiganj, a small town in West Bengal where the days move slowly and "neural network" wasn't a phrase anyone used. For a long time I assumed AI was a locked door — a club for people with PhDs and heavy calculus textbooks. I don't have a computer science degree. I didn't even study mathematics in college.

Then one night I found a video of an AI painting like a human artist, and something clicked. I didn't understand the equations, but I understood the idea: patterns, logic, a machine learning to see the world the way we do. So I skipped the heavy math and went straight for the concepts — reading, watching, building — until I understood how neural networks pass information around, how LSTMs remember, how a model actually learns to talk.

I'm still that same person from Raiganj. No fancy degree — just an endless hunger to keep building things that work.

AI & Data Science

Breaking complex architectures into functional blueprints for real-world impact.

Education

Self-study and specialized certification over rote memorization — depth came from doing.

Practical Work

Every experiment documented and measured — usability and clear results over theory alone.

2021 — 2023

B.Sc. (General), Biology

Siliguri College, Siliguri, India

Coursework in Botany, Zoology and Chemistry — a scientific-method foundation that still shapes how I frame problems and read results.

Feb 2025 — May 2026

Data Science & AI/ML Certification Trainee

Intellipaat / iHUB DivyaSampark, IIT Roorkee · Roorkee, India

A 16-month hands-on specialization covering statistics, EDA, supervised and unsupervised learning, NLP, deep learning and Generative AI. Built 25+ end-to-end ML workflows — including neural networks implemented from scratch before moving to TensorFlow/Keras.

May 2026 — Present

Current Focus & Experiments

Exploring new AI tools, vibe-coding, building websites, creating videos with AI, and building AI agents.

GenAIAgentsAI VideoWeb Apps

No CS degree. No calculus. 25+ ML systems shipped anyway.

Learned by building, not by theorem.

AI Job-Lead Pipeline

Research Agent

AI Data Analyst Agent

RAG-Based Conversational AI

Sentiment Analysis on Amazon Fine Food Reviews

Netflix/TMDB Movie Recommendation System

CS:GO Match Outcome Prediction

Handwritten Digit Recognition

Census Income Classification

Heart Disease Risk Prediction

Flight Price Prediction

Bio-Signal Smoking Detection

EN → JP Neural Machine Translation

Image Classification Benchmark

HR Analytics — Hypothesis Testing

From curiosity to code.

B.Sc. (General), Biology

Data Science & AI/ML Certification Trainee

Current Focus & Experiments

What I reach for.

Generative AI & LLMs

Machine Learning

Deep Learning

Natural Language Processing

Data Analysis & Statistics

Programming & Tools

Let's build something.