M karthikesh

M karthikesh

AI/ML Engineer · Intelligent Systems Builder

Hyderabad, India


About

I build AI-powered systems that combine machine learning, intelligent decision-making, and backend engineering to solve practical problems. My work spans anomaly detection, applied NLP, model evaluation, and full-stack development. I’m particularly interested in reasoning systems, adaptive planning, and the engineering behind reliable, low-latency AI applications. My projects include a real-time API abuse and anomaly detection system combining statistical methods, deep learning, and streaming feature extraction, and a hybrid sentiment analysis system integrating transformer-based models with rule-based scoring. I’m also exploring ECRS, an experimental cognitive routing system for language-model reasoning, and LifeOS, a personal operating system for planning, learning, and long-term goals. I build AI systems that have to be fast and reliable, not just accurate. My flagship project is a real-time API abuse detector: rules, statistical models and a PyTorch autoencoder + LSTM behind a Redis feature pipeline, at under 15 ms average latency. I’m also researching ECRS, a framework that routes language-model queries by complexity, so easy questions get cheap reasoning and hard ones get bounded verification. Currently strengthening my foundations in computer science while building increasingly rigorous AI systems and preparing for GATE CS & DA. Open to connecting with AI/ML engineers, researchers, developers, and builders interested in intelligent systems, applied research, and ambitious engineering projects.

Skills
Java
JavaScript
TypeScript
SQL
React
FastAPI
Node.js
PyTorch
Scikit-learn
XGBoost
Hugging Face Transformers
NLP
RAG
AWS
Amazon Bedrock
Amazon OpenSearch
AWS Lambda
Supabase
PostgreSQL
Redis
Git
Linux
Data Structures & Algorithms
Next.js
Google Cloud
Gemini
TensorFlow
System Design
Experience

Backend Developer – Payments & Order Management

TROVX PRIVATE LIMITED

Jul 2026 to Present
  • Integrated payment gateway workflows for HungerTap, covering payment initiation, verification, transaction status tracking, and refund handling.
  • Implemented order hash-key generation and verification to strengthen order integrity using the cryptographic mechanism configured in the system.
  • Implemented atomic stock deduction using PostgreSQL transactions and row-level locking to prevent overselling during concurrent orders.
  • Developed database-driven order and payment workflows using Supabase PostgreSQL RPC functions and triggers.
  • Implemented safeguards for failed, delayed, and duplicate payment events to improve transaction reliability and reduce inconsistent order states.
  • Strengthened inventory and payment consistency by revalidating stock and handling refunds when items became unavailable during payment processing.
Projects

Real-Time API Abuse & Anomaly Detection System

Architected a 3-layer real-time defense framework combining deterministic rules, statistical EWMA/Isolation Forest, and an unsupervised PyTorch Autoencoder + LSTM network for detecting previously unseen behavioral anomalies. Engineered an in-memory Redis feature extraction pipeline processing 30+ dynamic attributes, achieving <15ms average latency (p95 <22ms) and 91.3% precision on simulated attack benchmarks. Developed an interactive dashboard for real-time traffic monitoring, confusion matrix tracking, and live threat telemetry.

Python
FastAPI
PyTorch
XGBoost
Redis

Hybrid Sentiment Analysis System

Engineered a hybrid NLP architecture integrating pretrained transformer RoBERTa with VADER rule-based sentiment scoring using confidence-aware dynamic weighting. Implemented test-time augmentation with text normalization variants to improve robustness on ambiguous samples. Achieved 96.4% accuracy and 0.95 F1-score on the IMDB benchmark dataset, outperforming the standalone RoBERTa baseline by 4.2%; deployed a Telegram bot interface for real-time scoring.

Python
RoBERTa
Hugging Face Transformers
VADER
NLTK
spaCy

LifeOS — AI-Native Personal Productivity and Life Management Platform

A personal productivity platform designed to unify academic planning, GATE and DSA preparation, project tracking, career goals, and personal wellbeing. Its documented target design combines constraint-aware planning with AI-assisted recommendations, aiming to suggest realistic next actions while accounting for deadlines, available time, and recovery needs. Some capabilities remain planned or unverified, so showcase only those you can demonstrate.

React Native
Expo
TypeScript
Constraint-Based Scheduling
Supabase
PostgreSQL
Supabase Auth
Row Level Security (RLS)
TypeScript
Deterministic Planning Engine
Zod

Evolutionary Cognitive Routing System for Compute-Efficient Language Model Reasoning

Designed and developed ECRS, a complexity-aware framework for improving language model reasoning through adaptive computation and bounded verification. The system routes queries based on estimated complexity, using lightweight reasoning for simpler tasks and retrieval-augmented thought graphs or bounded Solver–Critic debate for more complex problems. It incorporates confidence aggregation, calibration, human-review escalation, and adaptive memory management to explore the trade-offs between reasoning accuracy, reliability, and computational cost. Benchmark Evaluation: Evaluated across 2,500 instances spanning five benchmarks: SQuAD v2 (question answering), HotpotQA (multi-hop reasoning), GSM8K (mathematical reasoning), StrategyQA (strategic reasoning), and TruthfulQA (truthfulness).

Python
Qwen2.5-3B
Llama 3.1 8B
Ollama
Retrieval-Augmented Generation (RAG)
Multi-Agent Reasoning
Confidence Calibration
M karthikesh · DevFest Hyderabad 2026