karthik = {
"role" : "ML Engineer · Data Scientist · Researcher",
"currently" : "@ Logical Box",
"published" : "5× IEEE Author across NLP, CV, sensors & GenAI",
"interests" : [
"Causal Inference & Experimentation",
"Video Understanding & Temporal Modeling",
"RAG & Agentic LLM Systems",
"Time Series & Predictive Maintenance",
"Whatever hard problem lands next 🤷"
],
"currently_building" : "Temporal video AI + causal user modeling research",
"ask_me_about" : "why my RNN underperformed a simple baseline, and what that actually meant",
"fun_fact" : "Ranked 354 / 75,000+ in Amazon ML Challenge"
}| 🎥 Temporal Video Understanding | 🎬 Causal OTT User Modeling |
|---|---|
| Frame-pooling vs. CNN+LSTM/GRU vs. CLIP zero-shot, head to head on UCF101 | Predictive + causal + policy modeling on MovieLens 25M |
| 96.11% Top-1 (simple avg-pool) — the RNN actually underperformed it, investigated down to the real cause instead of hidden | Hand-built IPW/AIPW causal estimators — AIPW recovers the true effect almost exactly |
| Every run + Grad-CAM explainability on real Kaggle GPU kernels, zero shortcuts | 78 tests · 0 leakage violations verified on the real 25M-row table |
 |
|
| 🛩️ AeroManual-RAG | 🕵️ FirmScout |
|---|---|
| Advanced RAG pipeline over 2,223 pages of FAA aerospace manuals | Multi-agent system that autonomously researches any company |
| Hybrid BM25 + FAISS · RRF Fusion · BGE Reranking | Finds AUM, deals & leadership — with confidence scoring and audit trails |
| RAGAS Answer Relevancy: 0.875 ✅ | Structured extraction, not just a chat wrapper |
|
|
| ⚙️ Predictive Maintenance | 💹 Financial Sentiment API |
|---|---|
| End-to-end RUL prediction for NASA turbofan engines | Fine-tuned FinBERT sentiment API |
| LSTM · TFT · Autoencoder Anomaly Detection | 88.57% accuracy on Financial PhraseBank |
| LSTM RMSE: 15.42 · R² = 0.862 ✅ | SHAP Explainability · FastAPI Risk API |
|
|
15 public repos total — see them all →
Languages
ML / DL
Causal Inference & Vision
GenAI / RAG / Agents
Deploy / MLOps
From fine-tuning one BERT model to a 5-domain research portfolio.
timeline
title
Apr 2025 : NLP foundations — BERT + SNLI fine-tuning
Nov 2025 : Applied ML — mental-health signal detection : recommender systems
Feb 2026 : Finance & explainability — Basel II credit risk (PyTorch, SHAP) : FinBERT sentiment
May 2026 : Rapid diversification — computer vision : multi-agent systems : RAG (LangChain, FAISS) : time-series (LSTM, TFT)
Aug 2026 : Research depth — causal inference (EconML, DoWhy) : video understanding (CLIP, Grad-CAM) : quant/HFT modeling : agentic RAG benchmarks
"If it's not published, did it even happen?"
- 📌 Machine Condition Monitoring Using TENG-Based Vibration Sensor — KST 2026 · DOI
- 📌 Multimodal Content Moderation (MemeShield-1K) — IEEE Transactions (submitted)
- 📌 +4 more IEEE conference papers — Google Scholar →
🥇 Amazon ML Challenge — 354th / 75,000+ | 📊 NPTEL Top 1% in LLMs | 📝 JEE Top 2% among 1M+