AI / ML Engineer · Agentic Systems · LangGraph · RAG · FastAPI
B.Tech CSE '28 @ AKGEC · Organizer & Mentor @ MLCoE · RAGFlow contributor
harsh_raj = {
"role" : "AI/ML Engineer · B.Tech CSE @ AKGEC (2028)",
"focus" : ["Agentic AI", "LangGraph / ReAct", "RAG Systems", "FastAPI backends"],
"now" : "AI Agentic Engineering Intern @ AxioraPeople",
"community" : "Organizer & Mentor @ Machine Learning Centre of Excellence (MLCoE)",
"open_source": "RAGFlow contributor (infiniflow/ragflow)",
"available" : "Open to AI/ML & backend internship opportunities",
}I build AI systems end-to-end, from architecture to deployment: autonomous multi-agent backends, real-time data platforms and RAG pipelines. Every project below is live, not just a notebook.
- 🔧 Taking AgentOS from a working multi-agent backend to a production-grade v2
- 🌍 Contributing to open-source RAG / agent infrastructure (RAGFlow), working toward bug fixes and tests
- 💼 Building AI-powered hiring features and backend services as an AI Agentic Engineering Intern
|
Multi-agent AI backend with human-in-the-loop safety LangGraph ReAct agent with a Supervisor routing to 3 specialised sub-agents (Research, Coder, Vision), three-tier memory (Postgres checkpointing, ChromaDB RAG, episodic audit logs) and confirm / reject / edit approval for risky tool calls. Approval flow verified end to end against a live Gmail account: a real send executed only after explicit approval. Stack:
|
Real-time space weather intelligence platform Tracks 27,000+ satellites in real time on live NASA / NOAA APIs, modelling per-satellite damage probability (0–100) via orbital mechanics and NASA SHIELDOSE-2. L1 solar-wind pipeline delivers 45-minute early warnings for geomagnetic storms, with grid heatmaps, G1–G5 risk alerts and a 3D Earth risk visualisation. Stack:
|
|
Agentic conversational data analyst LangGraph agent that turns plain-English questions into SQL / Pandas over DuckDB's OLAP engine, then surfaces hidden data issues with 3-layer anomaly detection (IQR, Z-Score, Isolation Forest), auto-generated Plotly charts and streamed AI summaries that explain its own reasoning. Stack:
|
Autonomous terrain navigation system Segmentation model on a Meta AI DINOv2 ViT-S14 backbone with a custom decoder built from scratch, reaching 98%+ validation accuracy on a 2GB+ dataset across 10 terrain classes. CLAHE preprocessing lifted low-visibility detection by 30%. Stack:
|
|
RAG-powered multi-service healthcare assistant 7 independent FastAPI microservices: disease prediction, drug-condition classifiers, food / medicine barcode lookup and a FAISS + Groq RAG chatbot (MediMind). Deployed on Render and consumed by a React web app and a Flutter mobile app. Stack:
|
LangGraph multi-agent content pipeline Research → Planning → Writing agent workflow with Tavily web search and Groq inference that turns a single topic into a structured, SEO-friendly blog post, fully automated end to end. Stack:
|
flowchart LR
UI[React UI] -->|SSE stream| API[FastAPI]
API --> SUP[Supervisor Agent]
SUP --> RES[Research Agent]
SUP --> COD[Coder Agent]
SUP --> VIS[Vision Agent]
RES --> TOOLS[Tools: web search, code interpreter, Gmail, Calendar, vision]
COD --> TOOLS
VIS --> TOOLS
TOOLS --> RISK{Risky action?}
RISK -->|yes| HITL[Human approval: confirm / reject / edit]
HITL --> TOOLS
API --> PG[(Postgres: checkpoints + episodic logs)]
API --> CHROMA[(ChromaDB: RAG memory)]
Design choices worth noting: Postgres (not Redis) for checkpointing so long-pending approvals survive restarts without TTL risk, SSE instead of WebSockets for one-way streaming of agent events, and a max-step guard that auto-terminates runaway agent loops.
| Area | What I did | Result |
|---|---|---|
| Agent safety | Human-in-the-loop approval (confirm / reject / edit) with a full audit trail in Postgres | Risky actions never run unapproved |
| Real-time systems | NASA / NOAA telemetry pipeline with orbital-mechanics damage modelling | 27,000+ satellites, 45-min early warnings |
| Computer vision | DINOv2 ViT backbone + custom decoder + CLAHE preprocessing | 98%+ val. accuracy, +30% low-visibility detection |
| ML engineering | Fixed a Random Forest data leak; built a KNN engine for 144,000 records | +63% real-world accuracy, no more memory crashes |
| MLOps | Automated 3 preprocessing pipelines, containerised deployments | 40% faster deployment, 15% faster processing |
| Project | Contribution | Status |
|---|---|---|
| RAGFlow | #20548: docs update adding a provider to the tested OpenAI-API-compatible model list | ✅ Merged |
Next up: bug fixes and tests in the RAG / agent space.
- AI Agentic Engineering Intern, AxioraPeople (Sep 2026 – Present, Remote): backend architecture, AI-powered hiring features (resume screening, AI video interviews) and containerised deployment
- Machine Learning Engineer, AKGEC MLCoE (Aug 2024 – Nov 2024): KNN engine for 144,000 records, Random Forest data-leak fix, pipeline automation
- Organizer & Mentor, AKGEC MLCoE (2025 – Present): mentoring juniors in applied ML / AI; led CONVOLUTE'26, a datathon with 1,344+ applicants
- 🎓 Oracle Agentic AI Certified Foundations Associate · Oracle Cloud Infrastructure Foundations Associate
- 🎓 Google DeepMind AI Research Foundations (Capstone) · Anthropic AI Fluency: Framework & Foundations
- 🎓 IBM SkillsBuild: Unleashing the Power of AI Agents · CampusX: FastAPI, Advanced RAG, Agentic AI Development
- 🥇 Amazon ML Challenge 2026: ranked under 700 among 100,000+ registrations (72-hour business entity resolution hackathon)
- 🥇 DeepThought CultureTech Fellowship: ranked #1 among shortlisted candidates
- 💻 450+ DSA problems solved · LeetCode contest rating 1575
- Agent evaluation and reliability: traces, evals, guardrails
- Advanced RAG: hybrid retrieval, re-ranking, evaluation
- Open-source contributions in RAG / agent infrastructure
- Production hardening for multi-agent backends: auth, multi-tenant OAuth, observability
⭐ If you like my work, consider starring my repositories.
"I don't build prototypes. I build things that ship."