Need an AI/LLM system integrated?
I can design RAG pipelines, dense embeddings, vector search (FAISS), custom model inference, prompt guardrails, and model evaluation.
RAG • Embeddings • FAISS • Evaluation
Available for Engineering & Product Roles
Atharv Pawar
Software & AI Engineer
I build and ship software end-to-end — from architecture and backend systems to AI/ML integration, automation, testing and deployment.
Technical generalist with hands-on experience across AI/ML, full-stack development, backend systems, APIs, automation, databases, security and deployment. I specialize in turning ambiguous technical problems into working products.
AI / ML Systems
Backend & APIs
Full-Stack Architecture
Automation & Pipelines
Real-Time Streaming
Security & Red Teaming
Rapid MVP Development
verified_stack
Hands-on experience
Python
PyTorch
Node.js
Express
React
Next.js
C++17
PostgreSQL
MongoDB
DockerCore Engineering Mindset
“Give me a problem. I'll figure out the stack.”
I don't limit myself to a single framework or layer of the stack. I start with the problem, understand the constraints, determine the technical approach, build the system, integrate the required services, test it, deploy it and iterate.
Engineering Impact
What I Can Take Ownership Of
Concrete problems I can take from ambiguity to production, freeing engineering leads and founders to focus on higher-level strategy.
Need a backend or API?
I can design REST APIs, backend services, JWT authentication, schema modeling, error handling, and serverless deployments.
Node.js • Express • REST • AuthNeed an end-to-end application or MVP?
I can take a concept from system requirements and architecture through frontend UI, backend logic, and production deployment.
Architecture • Full Stack • Rapid MVPNeed automation or data pipelines?
I can build automated document ingestion pipelines, PDF extraction, dataset deduplication, cleaning, and batch workflows.
Pipelines • Document ETL • ScrapingNeed real-time streaming or voice AI?
I can engineer bidirectional WebSocket streams with audio chunking, Voice Activity Detection (VAD), STT, and low-latency TTS.
WebSockets • VAD • Streaming • AudioNeed an internal dashboard or admin portal?
I can build responsive interfaces, authentication layers, role-based workflows, and data visualizations for operational tools.
Admin Portals • CRUD • VisualizationsExisting codebase broken or sluggish?
I can investigate system bottlenecks, debug asynchronous edge cases, refactor messy logic, and stabilize application behavior.
Debugging • Profiling • RefactoringNeed architectural direction under uncertainty?
I can research options, build rapid proof-of-concepts, evaluate trade-offs, and implement the decided approach cleanly.
System Design • Trade-offs • PoCDeep-Dive Architectures
Technical Case Studies
Verified architectures showing problem decomposition, systems engineering, AI/ML implementation, and technical decisions.
Research & Prototype
AI Security • Autonomous Red Teaming
AegisSwarm: Autonomous Multi-Agent AI Red-Teaming & Threat Simulation Framework
An event-driven multi-agent orchestration architecture designed to simulate, execute, and evaluate adversarial attacks against LLM agents, tools, memory stores, and RAG pipelines.
System Architecture & Execution Pipeline
01
Threat Model
Jailbreaks, Injections, Poisoning
→
02
Event Bus
Async Message Broker
→
03
HTN / ToT Planner
Multi-Step Attack Chains
→
04
Attacker Agents
Tool / Memory Hijackers
→
05
Security Benchmark
ASB • AgentDojo • Reports
The Problem
Modern LLM agents with access to tools, RAG memory, and external APIs are vulnerable to multi-turn exploits such as indirect prompt injection, long-term memory poisoning, and tool hijacking. Static one-shot benchmarks fail to capture adaptive, multi-step agent attack chains.
Threat Model Coverage
- Indirect Prompt Injection: Embedding malicious payload instructions into untrusted external data and tool returns.
- Memory & RAG Poisoning: Injecting deceptive context into persistent episodic memory or vector retrieval indexes.
- Tool Hijacking & Data Exfiltration: Tricking autonomous agents into invoking unauthorized privileged functions.
Engineering & Architecture Decisions
- Microkernel Architecture: Core runtime handles async messaging and state while attack/defense modules load dynamically via plugin interfaces.
- Planning Search (HTN + Tree-of-Thoughts): Hierarchical Task Networks decompose high-level exploit goals into tactical sub-tasks, while ToT explores and scores candidate attack branches.
- Standardized Benchmark Alignment: Designed to evaluate against state-of-the-art security suites (Agent Security Bench [ASB], AgentDojo, and HarmBench).
Core Technologies:
Python
Multi-Agent Systems
Async Event Bus
HTN Planning
Tree-of-Thoughts
LLM Red Teaming
Implemented & Validated
Scientific RAG • Custom PyTorch Transformer
GrindHaus AI: Grounded Scientific Document RAG & Custom PyTorch Transformer Pipeline
A single-machine retrieval-augmented generation system and 20-layer custom GPT-style Transformer engineered for grounded, hallucination-resistant sports science reasoning.
End-to-End RAG & Custom Transformer Architecture
01
Research Papers
PDF Extraction & Cleaning
→
02
Chunking & Embeddings
all-MiniLM-L6-v2 Dense Vectors
→
03
Native C++ FAISS
Vector Similarity Search
→
04
Intent Router
Expert vs. Companion
→
05
20-Layer Transformer
RoPE • SwiGLU • KV Cache
The Problem
General-purpose LLMs frequently generate ungrounded or contradictory fitness and nutritional advice. The goal was to build a system providing strictly grounded reasoning backed by peer-reviewed sports science literature with low-latency local execution.
Data & RAG Ingestion Pipeline
- Automated Document Pipeline: Implemented drag-and-drop PDF extraction, sliding-window chunking, and dense embedding generation using
all-MiniLM-L6-v2. - C++ FAISS Vector Search: Handled high-throughput nearest-neighbor similarity search natively via FAISS to minimize query latency on Windows.
- Curated Instruction Dataset: Built and deduplicated an 800-sample domain dataset (400 Expert Mode + 400 Companion Mode) with zero repeated assistant responses.
Custom PyTorch Transformer Architecture
- Model Specifications:
hidden_size=1536,num_layers=20,num_heads=12,intermediate_size=4096,vocab_size=32000,max_seq_len=2048. - Modern Architectural Primitives: Rotary Positional Embeddings (RoPE), SwiGLU non-linear activations, RMSNorm, and KV-cache inference optimization.
- Intent-Based Routing: Dynamic inference router dispatching scientific queries to Expert mode and conversational queries to Companion mode.
Core Technologies:
Python
PyTorch (bf16)
FAISS (C++)
SentenceTransformers
Custom Transformer
RAG
Production Monorepo
Founder • Systems & AI Product Engineering
REDAESTH: Full-Stack Monorepo Platform with Local C++ Companion Engine
A production-ready monorepo combining a responsive React client, modular Node.js/Express REST API (/api/v1), and a standalone C++ coaching engine with persistent local memory.
Platform Monorepo & Service Architecture
01
React Client
Styled Components • Framer • Lenis
→
02
Node/Express API
Versioned /api/v1 Endpoints
→
03
Local C++ Engine
C++17 Coaching Engine
→
04
Persistent Context
Local memory.json Context Store
→
05
Cloud Deployment
Vercel Serverless • Auth • CDN
Independent Technical Ownership
As founder and lead engineer, I architected the complete platform ecosystem — establishing codebase structure, backend routing, persistent context systems, client state management, and deployment pipelines.
Modular Backend & API Architecture
- Versioned REST API: Built structured `/api/v1` routes for authentication (signup/login), user profiles, chat, and community post/like/comment workflows.
- Serverless & Container Delivery: Configured dual runtime support — standard Node server for local testing and Vercel serverless functions for zero-maintenance production deployment.
Local AI Companion & Systems Engineering
- Zero-Cloud-Cost Local Companion: Designed a local conversational coaching system updating persistent JSON context memory without paying recurring token fees to external APIs.
- C++ Coaching Integration: Built an optional standalone C++17 engine (
grind_engine.exe) for deterministic local reasoning and sub-millisecond responses. - Always-On Simulator: Background health checking and streak monitoring for user workout and hydration consistency.
Core Technologies:
Node.js
Express
React
C++17
REST APIs
Vercel Serverless
Technical Breadth
Engineered Capabilities
A structured breakdown of my hands-on technical competencies across architecture, AI/ML, backend engineering, data, and infrastructure.
01
Architecture & Product Engineering
- Requirements analysis & technical scoping
- System architecture & interface contracts
- Database schema & data model design
- REST API architecture & versioning
- Rapid MVP development & prototyping
- Pragmatic technical decision-making & trade-offs
02
AI & Machine Learning
- LLM application development & prompt design
- RAG architectures & document chunking
- Dense embeddings & FAISS vector databases
- Custom Transformer modeling (PyTorch, RoPE, SwiGLU)
- Multi-agent orchestration & HTN planning
- AI security, prompt injection & red teaming
03
Application Engineering
- React & Next.js component architectures
- JavaScript (ES6+) & TypeScript fundamentals
- Node.js & Express server development
- Python scripting & machine learning control layers
- RESTful API design & integration
- JWT authentication & session security
04
Real-Time & Automation
- WebSocket bidirectional audio/data streaming
- Voice AI pipelines (VAD, STT, LLM, TTS)
- Automated document extraction & processing
- Data scraping & ETL workflows
- Third-party API integrations & webhooks
- Parallelized async pipeline optimization
05
Data & Persistence
- PostgreSQL relational database design
- MongoDB document modeling
- SQLite & Android Room Database
- FAISS native vector stores
- Persistent context & cache management
- CRUD operations & transaction management
06
Infrastructure & DevOps
- Git version control & GitHub workflows
- Docker containerization basics
- Linux / Bash command-line tooling
- Vercel cloud & serverless deployment
- Environment configuration & secrets management
- CI/CD build and test automation
07
Security & Reliability
- Authentication & role-based authorization
- Input validation & sanitization
- API security & rate limit strategies
- LLM prompt injection defense research
- Systematic debugging & root cause analysis
- Unit testing & validation commands
Verified Software
Selected Engineering Projects
A showcase of built and deployed systems demonstrating full-stack, mobile, and async engineering.
Real-Time Voice AI Pipeline
Low-latency bidirectional voice system connecting browser audio to backend via WebSockets with energy-based VAD, STT, streaming LLM, and TTS synthesis.
Engineering Highlights:
Audio chunking at ~20ms, non-blocking parallelized pipeline stages, and modular architecture allowing seamless model substitution.
Therapist Platform / Client Web Application
Production-level full-stack responsive web application engineered for a practicing therapist with custom layout systems, accessible UI, and sub-second load times.
Engineering Highlights:
Custom typography scaling, micro-interactions, responsive mobile-first architecture, and automated cloud delivery on Vercel.
ExpenJar — Offline-First Android Expense App
Native Android application in Java using SQLite via Room Database for persistent offline transaction tracking, spending analytics, and category breakdowns.
Engineering Highlights:
MVVM architectural pattern, transactional ACID database handling, reactive UI state updates, and zero-network local storage reliability.
Live Weather & Forecast Dashboard
Asynchronous API-integrated weather application featuring 5-day predictive forecasts, dynamic DOM manipulation, and defensive error state management.
Engineering Highlights:
Non-blocking asynchronous fetching with async/await, debounced search queries, and real-time DOM element reconstruction.
Execution Framework
How I Work
A repeatable 7-step engineering methodology that eliminates ambiguity, minimizes technical risk, and ships dependable software.
01
Understand
Deconstruct requirements, operational constraints, user intent, and the exact root problem before writing code.
02
Architect
Design data flow, interface schemas, database models, and choose the simplest appropriate technology stack.
03
Build
Implement core functionality with modular components, clear typing, and defensible error boundaries.
04
Integrate
Connect REST APIs, persistent databases, AI models, vector stores, and external third-party services seamlessly.
05
Test
Systematically debug edge cases, benchmark latency, validate inputs, and ensure cross-platform stability.
06
Deploy
Package the solution into containers or serverless runtimes with reproducible build scripts and clean environment variables.
07
Iterate
Measure real-world performance, profile bottlenecks, gather feedback, and continuously refine the product.
Background
About Me
Who I Am
I am a Computer Science Engineering undergraduate at Government College of Engineering, Chandrapur, and the founder of REDAESTH Pvt. Ltd.
I operate as an end-to-end builder — comfortable working across machine learning pipelines, backend services, database schemas, real-time audio streams, and frontend user interfaces.
Engineering Focus
I specialize in taking ambiguous, messy technical challenges and engineering clean, grounded solutions that actually work.
Whether it's building a scientific RAG pipeline with custom vector indexing, orchestrating multi-agent AI security simulations, or architecting a production monorepo, I take end-to-end technical responsibility.
Core Strengths
- End-to-End Ownership: From system design to deployment
- AI & Systems Focus: RAG, PyTorch, C++, and WebSockets
- Founder Mindset: High autonomy, pragmatism, and execution speed
- Remote Work Ready: Structured communication and clean documentation
Let's Connect
Get in Touch
Open to Software & AI Engineering roles, technical collaborations, and startup projects. Let's discuss what we can build together.