# Milan Soni — AI Engineer & Full Stack Developer

> AI Engineer and Full Stack Developer building production RAG pipelines, multi-agent orchestration systems, and multi-provider LLM infrastructure. Shipped code to a 46.9k★ open-source AI gateway (3 PRs merged, 21,000+ tests), built a 6-agent document intelligence platform that won SIH 2023 National (Ministry of Coal, 44,000+ teams), and deployed $0/month production systems using free-tier AI providers. Scopus-indexed researcher. B.Tech CSE '26.

## Profile

- **Name**: Milan Soni
- **Role**: AI Engineer & Full Stack Developer
- **Location**: Churu, Rajasthan, India
- **Education**: B.Tech CSE, Global Institute of Technology, Jaipur (CGPA: 8.10, graduating 2026)
- **Email**: milansoni96946@gmail.com
- **GitHub**: https://github.com/Iammilansoni
- **LinkedIn**: https://www.linkedin.com/in/sonimilan/
- **Medium**: https://medium.com/@milansoni96946
- **Portfolio**: https://milansoni.vercel.app

## Core Expertise

- **AI / LLM Engineering**: Production RAG pipelines with hybrid search (vector + BM25), multi-agent orchestration (6 agents), cross-encoder reranking, evaluation harnesses
- **Full Stack**: React 19, Next.js 16, TanStack Start, FastAPI, Clerk auth, enterprise RBAC
- **Data & Infrastructure**: PostgreSQL + pgvector, Redis HNSW indexes, Docker, $0/month free-tier deployments
- **Systems Thinking**: Algorithm design, scalable backend architecture, multi-provider AI orchestration

## Awards & Recognition

- **SIH 2023 National Winner** — Top 1% out of 44,000+ teams, recognized by Coal India Limited & CMPDI
- **Scopus-Indexed Researcher** — Peer-reviewed paper on hybrid attention-based temporal modeling (PICET-2026, IET Conference Proceedings)

## Projects

### MiningNiti — AI Document Intelligence for Mining

> SIH 2023 National Winner | Recognized by Coal India Limited & CMPDI

A full-stack AI platform combining a multi-agent AI pipeline (6 specialized agents across 4 AI providers) with production-grade RAG (hybrid search + cross-encoder reranking) and real-time compliance auditing. Built for the Ministry of Coal.

- **Agents**: Classifier (Groq/Llama 3.3), Safety Analyzer (Mistral/Magistral), Entity Extractor (Cerebras/GPT-OSS-120B), Summarizer (Cerebras), Compliance Auditor (Gemini), Orchestrator
- **RAG Pipeline**: Hybrid search (pgvector cosine + pg_trgm BM25) → Reciprocal Rank Fusion → ms-marco-MiniLM-L-6-v2 cross-encoder reranking → Top-5 chunks → LLM generation
- **Tech**: Next.js 16, React 19, FastAPI, PostgreSQL + pgvector, Supabase, Upstash Redis, Clerk Auth, Groq, Cerebras, Mistral, Gemini, Docker
- **Metrics**: 6 AI agents, 4 providers, $0/month infrastructure cost
- **Impact**: Won SIH 2023 National Finale. Reduced manual compliance analysis by 90%. Pipeline completion rate: 98.7%. End-to-end latency: 3.2s (down from 14s sequential).
- **GitHub**: https://github.com/Iammilansoni/MiningNiti
- **Demo**: https://miningniti.vercel.app/

### HATF Early Warning — Research Paper to Shipped Product

> Scopus-indexed publication (PiCET-2026, IET Conference Proceedings) implemented as a working AI product

An end-to-end early-warning system for student dropout, built from Milan's co-authored paper on the Hybrid Attention Temporal Framework. Predicts risk from week 2 of an 8-week course, explains every score in behaviours an advisor can verify, quantifies its own uncertainty, and audits its own fairness.

- **Model**: 59,951-parameter HATF — multi-scale causal convolutions (kernels 1/3/7) → unidirectional LSTM → masked multi-head temporal attention → MC-dropout head (30 passes). One checkpoint serves every prediction week, trains on a laptop CPU.
- **Leakage control**: causality is structural — left-padded convolutions and a unidirectional LSTM make future weeks physically unreachable, proved by tests that overwrite masked weeks with noise 50×.
- **Explanations**: occlusion-based sensitivity, shown as raising risk only when model sensitivity and the student's deviation from the cohort agree, so an explanation cannot contradict itself.
- **Results** (held-out test students, synthetic data): ROC-AUC 0.789 at week 2 → 0.876 at week 8; pooled 0.845 (95% CI 0.751–0.924); expected calibration error 0.038; 28.4% escalated for human review; counterfactual invariance exactly 0.000000; 110 tests.
- **Intellectual honesty**: eleven models trained under identical conditions and HATF finishes last on the demo cohort — the table is published anyway. It ships because at statistically indistinguishable accuracy it is the only model that also produces the attention, uncertainty and window-usage the explanation layer needs. The fairness regulariser is reported as a measured no-op (0.0000 change), and attention came out flat so the system says so rather than faking importance.
- **Tech**: PyTorch (CPU), FastAPI, Next.js 15, React 19, TypeScript, Tailwind v4, pandas, scikit-learn, Streamlit, pytest, Docker, uv
- **Deployment**: FastAPI on Render (Docker), Next.js dashboard on Vercel
- **GitHub**: https://github.com/Iammilansoni/hatf-lms-early-warning-poc
- **Demo**: https://hatf-lms-early-warning-poc.vercel.app/
- **Note**: the paper reports F1 94.2% / AUC 96.1% on 7,935 real students; this implementation runs on synthetic data and reports its own independently measured numbers. The two are not comparable.

### NLPForge — NLP Dataset Generator & Semantic Search

> Enterprise AI NLP Platform (built at nTheta Works)

An enterprise-grade platform bridging natural language and API testing. Two-stage retrieval pipeline (KNN vector similarity + FlashRank neural re-ranking) produces structured, executable API test cases.

- **Pipeline**: Ollama embeddings (nomic-embed-text) → Redis Stack HNSW indexes → FlashRank cross-encoder (ms-marco-MiniLM-L-12-v2) reranking
- **Tech**: Next.js 16, FastAPI, SQLAlchemy 2.0, PostgreSQL 15, Redis Stack 7.2, FlashRank, Ollama, Docker
- **Metrics**: 8 LLM providers, 15+ embedding models, 40% accuracy improvement, 60% QA effort reduction
- **GitHub**: https://github.com/Iammilansoni/NLPFT-2

### FinSageAI360 — AI Financial Intelligence

> Built at OBG Outsourcing

A multi-tenant SaaS that ingests accounting feeds, normalises them, and surfaces AI-generated cashflow, risk, and anomaly insights.

- **Tech**: Next.js, Node.js, MongoDB, Prisma, JWT, AI Analytics
- **Metrics**: +45% faster report cycles, +30% operational efficiency, enterprise-grade RBAC

### SmartLearnX — AI-Powered LMS

> Freelance Project

An adaptive educational ecosystem integrating AI, ML, NLP, and modern full-stack web technologies.

- **ML Models**: Logistic Regression dropout prediction (91.4% accuracy), Random Forest academic forecasting (0.89 R²)
- **NLP**: BERT quiz generation, spaCy chatbot assistant
- **Tech**: React, TypeScript, Node.js, FastAPI, MongoDB, Redis, Docker, Machine Learning

## Experience

| Role | Company | Period | Key Impact |
|------|---------|--------|------------|
| Open Source Contributor | OmniRoute (46.9k★) | Jul 2026 | 3 PRs shipped: HTTP 400 fix (schema adoption), provider filter (9/9 tests), Claude 5 Sonnet integration, 42-locale docs normalization |
| Full Stack Developer Intern | nTheta Works | Oct – Dec 2025 | Two-stage semantic retrieval pipeline: 40% accuracy improvement, 60% QA reduction |
| AI & Full Stack Developer | Freelance | Jul – Aug 2025 | ML dropout prediction (91.4%), performance forecasting (R²=0.89), 24/7 NLP chatbot |
| Full Stack Developer Intern | OBG Outsourcing | May – Jul 2025 | Led FinSageAI360: +45% report speed, +30% operational efficiency |
| Software Developer Intern | Om Logistics | Jun – Aug 2024 | LangChain + FAISS: 70% latency reduction across 10K+ docs |

## Tech Stack

- **Languages**: JavaScript (ES6+), TypeScript, Python, C++
- **Frontend**: React 19, Next.js 16, TanStack Start/Router/Query, Tailwind CSS v4, Framer Motion, Recharts, Radix UI / shadcn, Three.js
- **Backend**: FastAPI 0.128, Node.js, Express.js, SQLAlchemy 2.0, Pydantic v2, JWT, Clerk Auth, RBAC, Microservices
- **AI / ML**: LangChain, LangGraph, RAG Pipelines, Hybrid Search (Vector + BM25), Cross-Encoder Reranking, FlashRank, Ollama, AI Agents, pgvector
- **LLM Providers**: Groq (Llama 3.3), Cerebras (GPT-OSS-120B), Mistral (Magistral), Google Gemini, OpenAI, Anthropic, DeepSeek, HuggingFace
- **Databases**: PostgreSQL + pgvector, Supabase, MongoDB, Redis Stack 7.2 (HNSW), Upstash Redis, Prisma ORM
- **Cloud / DevOps**: Vercel, Nitro, HuggingFace Spaces, Docker Compose, GitHub Actions CI/CD, Linux, Git

## Blog Articles

### RSC + Streaming LLMs: Designing a Zero-Latency AI Dashboard with Next.js Server Actions
React Server Components + chunked streaming for AI dashboards. TTFB dropped from 3.2s to 120ms. Client JS reduced by 40%. Covers Server Actions, Vercel AI SDK, progressive rendering with Framer Motion.

### The State Machine Paradigm: Why I Ditched Linear LLM Chains for LangGraph Multi-Agent Workflows
Migration from LangChain SequentialChain to LangGraph for MiningNiti's 6-agent pipeline. Pipeline completion: 72% → 98.7%. Latency: 14s → 3.2s via asyncio. Covers state management, conditional edges, error resilience.

### Stop Overpaying for Vector DBs: Building a Production-Ready Hybrid RAG Pipeline for $0/Month
Replaced Pinecone ($400/mo) with pgvector on Supabase free tier. Hybrid search (BM25 + cosine) with Reciprocal Rank Fusion. Cross-encoder reranking with FlashRank. <10ms queries on 100K vectors.

### How I Built a Production-Grade Multi-Agent AI System That Won a National Hackathon
Deep dive into MiningNiti's architecture: 4 agent specialists (Classifier, Safety Analyzer, Entity Extractor, Summarizer), async background processing, RAG-powered chat, Celery task queue.

## Research

### Hybrid Attention-Based Temporal Modeling for Early Dropout Prediction in Learning Management Systems
Peer-reviewed conference paper, **Scopus indexed**.

- **Authors**: Pradeep Jha, Manju Mathur, Abhay Purohit, **Milan Soni** (4th author), Avadhi Singhal, Abhilash Joshi — Department of CSE, Global Institute of Technology, Jaipur
- **Venue**: 8th Parul University International Conference on Engineering & Technology (PiCET-2026), 1–2 May 2026
- **Publication**: IET Conference Proceedings · Paper ID PU/PiCET26/COP/327 · Accepted, in press
- **Contribution**: HATF combines multi-scale temporal convolution, recurrent modelling, temporal attention for interpretability, Bayesian-style uncertainty, and fairness-aware evaluation. Evaluated on three real LMS datasets — 7,935 students across 45 courses — reporting F1 94.2% and AUC 96.1%.
- **Implementation**: shipped as a working product (see HATF Early Warning above), delivering one of the paper's own stated future-work items — a dashboard giving educators timely notifications.
- **Paper**: https://drive.google.com/file/d/11DTgnEqtFGIB-PpX-SKyheMCue5xRe-_/view?usp=sharing

### Deploying Agentic AI in Production
State machines over prompt engineering for production agents. Critiques ReAct loops (infinite loops, context exhaustion, unpredictable latency). Advocates LangGraph deterministic orchestration with Pydantic/Zod guardrails.

## Portfolio Website

This repository (milan-vision-labs) is a production-grade personal portfolio built with:
- **Framework**: TanStack Start (React 19 + Vite + Nitro)
- **Styling**: Tailwind CSS v4 with oklch "Aurora" design tokens, glassmorphic effects
- **3D**: Three.js neural network particle system (lazy-loaded)
- **AI Chat**: Gemini-powered chat assistant with portfolio context
- **Blog**: Markdown-based with TOC, syntax highlighting, Medium RSS integration
- **SEO**: JSON-LD structured data, Open Graph, sitemap, robots.txt
- **Deploy**: Vercel (auto-deploy from main branch)

## Contact

| Platform | Link |
|----------|------|
| Email | milansoni96946@gmail.com |
| LinkedIn | linkedin.com/in/sonimilan |
| GitHub | github.com/Iammilansoni |
| Medium | medium.com/@milansoni96946 |
| Portfolio | milansoni.vercel.app |