bhaskar = {
"role": "Machine Learning Engineer",
"location": "Mumbai",
"focus": ["Applied ML/DL", "GenAI Pipelines", "MLOps", "Edge Deployment"],
"currently": "Building AnimeGAN — GAN model from scratch that generates anime like images.",
"kaggle": "Top 2% finishes across multiple Playground Series",
"philosophy": "Understand the math. Ship the model. Repeat.",
}- 🔭 Building BareMetalML — a pure-NumPy ML library for first-principles understanding
- 🤖 Ex-AI Engineer Intern @ Arcitech AI and Ebizon Cloud
- 🏆 Kaggle Top 2% — consistently placed in the top bracket across tabular & LLM competitions
- 📐 I care deeply about why models work, not just that they work
Languages
ML / DL Frameworks
GenAI & LLM
MLOps & Cloud
An open-source ML library built from the ground up using only NumPy — no scikit-learn, no shortcuts.
Implementing every algorithm end-to-end with full mathematical transparency:
- ✅ Linear Regression — Normal Equation, Batch/SGD, Ridge (L2) regularization
- ✅ Logistic Regression — Sigmoid, Binary Cross-Entropy, One-vs-Rest multiclass
- ✅ Decision Trees — Gini Impurity / Entropy splitting, decision boundary visualization
- 🔄 Random Forest — Bootstrap Aggregation (Bagging) (in progress)
- 🏗️ Modular architecture: custom train/test split, StandardScaler, MinMaxScaler, MSE/R² metrics
Python · NumPy · Pandas · Matplotlib
🏢 Arcitech AI — AI Engineer Intern Jul 2025 – Dec 2025 · Mumbai
Speech Emotion Recognition (SER)
- Designed & trained a lightweight CNN from scratch on 4 benchmark datasets (RAVDESS, TESS, CREMA-D, SAVEE)
- Achieved ~88% accuracy across 7 emotion classes with a model under 50 MB / 2M parameters — edge-deployable
Production Audio Pipeline
- Ingested call recordings from Twilio; prototyped with OpenAI Whisper (LangChain), then migrated to AssemblyAI after benchmarking showed superior WER on noisy BPO audio
- Built on top: automated summarization, Named Entity Recognition (NER), PII redaction (phone, email, CC, SSN)
- Processed 100–150 calls/run averaging 15 min each with full compliance-safe output
12-KPI Analytics Engine
- Engineered 3 primary KPIs (voice intensity/emotion scoring, contextual acknowledgement, quantitative performance) + 9 derived KPIs powered by SER + NLP outputs
Deployment
- Deployed entire pipeline as a production FastAPI microservice with async request handling and audio chunking for parallelized AssemblyAI calls
- Sole AI contributor; led all R&D, model selection, and MLOps decisions over 5 months
🏢 Ebizon Cloud — AI Engineer Intern Mar 2025 – May 2025 · Remote (US)
- Built an end-to-end RAG pipeline over Oracle Cloud Infrastructure (OCI) documentation — deployed as an internal knowledge-assistance tool
- Benchmarked 5+ LLMs (GPT, Gemini, Mistral, Nemotron, DeepSeek V3) via OpenRouter against an 80–100 question evaluation set — DeepSeek V3 and Gemini led on accuracy and grounding
- Built retrieval layer using FAISS with OpenAI embeddings and fixed-token chunking
- Designed a hallucination-mitigation framework: confidence scoring + citation/source verification + answer cross-referencing
- Iteratively refined chunking strategy, prompt design, and model selection to reduce unsupported outputs
| Competition | Rank | Field | Percentile |
|---|---|---|---|
| Academic Success Classification | 47 / 2,684 | Tabular Classification | 🥇 Top 2% |
| Irrigation Need Prediction | 86 / 4,315 | Tabular Classification | 🥇 Top 2% |
| Calorie Expenditure Prediction | 150 / 4,316 | Tabular Regression | 🥈 Top 4% |
| LLM 20 Questions (Featured) | — / 832 | LLM Agent | 🤖 LLM Strategy |
Stack: XGBoost · LightGBM · CatBoost · Feature Engineering · Optuna · Model Ensembling
⚡ "The best model is the one that ships — built on the math you actually understand."