I'm a Backend Engineer passionate about the intersection of highly scalable distributed architectures and intelligent AI agent systems.
Currently, I'm working at Freecharge (Axis Bank), designing and optimizing backend systems for core Investment and Personal Finance Manager products that serve millions of digital banking users.
As an Aerospace Engineer turned Software Developer π, I bring a strong engineering foundation, a first-principles problem-solving mindset, and a deep interest in low-latency systems, machine learning, and multi-agent orchestration.
- π Currently working on: Scaling fintech transaction engines & exploring custom agent runtime frameworks.
- π± Learning & Researching: Distributed systems consensus, low-level LLM compilers, and high-throughput LLM serving.
- π¬ Ask me about: Node.js, Python, AWS Cloud & Serverless architecture, LLM inference, LangGraph, and RAG systems.
- β‘ Fun Fact: Transitioned from designing physical flight trajectories in aerospace to engineering digital data flows and AI agent pathways.
-
Maestro | Multi-Agent Research Workflow Engine
- Engineered a multi-agent orchestrator using LangGraph managing a 4-agent team (Supervisor, Researcher, Writer, Critic) with quality-gated revision loops.
- Streams real-time tokens and agent state transitions via Server-Sent Events (FastAPI / SSE).
- Tech: Python, LangGraph, Anthropic Claude API, BM25 Search, Tavily.
-
NexusAI | Stock Analysis & Research Copilot
- Developed a self-hosted financial research copilot executing custom LangGraph workflows for multi-market equities, forex, and crypto.
- Implemented a Local RAG pipeline (ChromaDB + sentence-transformers) over SEC filings and Yahoo Finance data.
- Optimized API performance using a Redis TTL cache with in-memory fallbacks for financial endpoints.
- Tech: Python, FastAPI, ChromaDB, Redis, Ollama, React, Docker Compose.
- FlashInfer | PyTorch LLM Inference Engine From Scratch
- Developed a lightweight LLM inference engine supporting GPT-2 and Llama-style models.
- Implemented KV-cache incremental decode, RMSNorm, RoPE, SwiGLU, and Top-K/Top-P sampling.
- Exposed streaming endpoints via FastAPI with SSE support and created CLI evaluation/benchmark suites.
- Tech: Python, PyTorch, FastAPI, Hugging Face Hub.
-
CogniPost | Autonomous AI Content Pipeline
- Created an automated daily agent posting system to X (Twitter) using Node.js and TypeScript.
- Features trend discovery (Google + X APIs), LLM-based fact generation, and a multi-stage scoring/guardrail evaluator.
- Tech: Node.js, TypeScript, OpenAI/Anthropic APIs, GitHub Actions.
-
Credit Card Fraud Detection | Streaming MLOps Pipeline
- Built an end-to-end credit card fraud detection system on AWS.
- Configured AWS SageMaker Feature Store and Amazon Kinesis for real-time feature aggregation (KDA SQL), triggering Lambda functions to update the online feature store and invoke an XGBoost inference endpoint.
- Tech: AWS (SageMaker, Kinesis, Lambda, CloudFormation), Apache Spark, XGBoost.
-
Hate Speech Detection | Modular NLP Pipeline
- Designed an end-to-end NLP classification pipeline to detect hate speech.
- Followed modular software design (Data Ingestion -> Transformation -> Model Trainer -> Pipeline), deploying a Streamlit web app.
- Tech: Python, Scikit-learn, Flask, Streamlit, Docker.
Frameworks, APIs & Data Tools: LangGraph | LangChain | ChromaDB | Ollama | Hugging Face | Apache Spark | Amazon Kinesis | Serverless Framework
βοΈ Custom Redesigned Portfolio Profile for Tathagat Sathe