███╗ ███╗██╗██╗ █████╗ ███╗ ██╗ ███████╗ ██████╗ ███╗ ██╗██╗
████╗ ████║██║██║ ██╔══██╗████╗ ██║ ██╔════╝██╔═══██╗████╗ ██║██║
██╔████╔██║██║██║ ███████║██╔██╗ ██║ ███████╗██║ ██║██╔██╗ ██║██║
██║╚██╔╝██║██║██║ ██╔══██║██║╚██╗██║ ╚════██║██║ ██║██║╚██╗██║██║
██║ ╚═╝ ██║██║███████╗██║ ██║██║ ╚████║ ███████║╚██████╔╝██║ ╚████║██║
╚═╝ ╚═╝╚═╝╚══════╝╚═╝ ╚═╝╚═╝ ╚═══╝ ╚══════╝ ╚═════╝ ╚═╝ ╚═══╝╚═╝
name: Milan Soni
role: AI Engineer · Backend & Full Stack Developer
location: Churu, Rajasthan, India
education: B.Tech CSE @ Global Institute of Technology, Jaipur — CGPA 8.10/10 ('26)
recognition:
- SIH 2023 National Winner — Ministry of Coal, top 1% of 44,000+ teams
- Scopus-indexed publication — PiCET-2026 (IET Conference Proceedings)
open_source: 6 PRs merged into OmniRoute — 53k★ AI gateway, 230+ providers, 21,000+ tests
open_to: [Full-time roles, Freelance projects, AI consulting]I build production RAG pipelines, multi-agent LLM systems, and the backends that hold them up — in Python and TypeScript. The through-line in my work is measurement: retrieval quality scored against a labelled golden set and enforced as a blocking CI gate, failure modes surfaced rather than swallowed, and limitations written down instead of rounded up.
|
Production RAG with hybrid search and cross-encoder reranking · multi-agent orchestration
(5 agents, no framework — hand-written on |
FastAPI · Node.js/Express · SQLAlchemy · Pydantic v2 · React 19 + Next.js 16 frontends · Clerk/JWT auth with RBAC · SSE streaming · REST API design |
PostgreSQL + pgvector (HNSW) · Redis vector search · Supabase · Docker Compose · GitHub Actions CI/CD · $0/month free-tier production deployments |
|
AI document intelligence & compliance for mining
5 specialized agents across 4 providers — 4 run on upload (classifier first, then three
concurrently), a 5th audits compliance on demand. No agent framework: orchestration is
hand-written on
🔗 Code · 🌐 Live · 📖 Spotlight |
A Scopus-indexed paper, shipped as a product
Early dropout prediction from week 2 of an 8-week course. A 59,951-parameter model — causal CNN → LSTM → masked temporal attention → MC-dropout — where causality is structural, not conventional. It explains every score, quantifies its own uncertainty (28.4% escalated to a human), and audits its own fairness. Eleven baselines were trained and the table is published even though the proposed model loses.
|
⚡ NLPForgeNL → executable API test cases
Two-stage retrieval: Ollama embeddings → Redis Stack HNSW KNN → FlashRank cross-encoder reranking, with LLM slot extraction across 8 providers and 15+ embedding models. +40% template-matching accuracy, −60% manual QA effort.
|
AI-powered Learning Management System
Adaptive LMS with dropout prediction (Logistic Regression, 91.4% accuracy) and performance forecasting (Random Forest, R² = 0.89) served from a FastAPI microservice, plus BERT quiz generation and a spaCy chatbot answering under 2s at load.
|
Two builds, four years apart. The Smart India Hackathon 2023 entry — a team prototype against the Ministry of Coal problem statement — won the National Finale and was recognized by Coal India Limited & CMPDI. This repository is not that codebase. It is an independent, ground-up rebuild started June 2025 and developed solo since, with none of the 2023 code carried over: a production system rather than a hackathon demo.
|
The 2023 prototype → single-pass LangChain RAG chatbot over FAISS, running locally. Today → 5 agents across 3 LLM providers with quota-aware failover, hybrid retrieval with cross-encoder reranking, a blocking retrieval-quality CI gate, Clerk JWT auth with SSRF and prompt-injection guards, and 36 REST endpoints live on free-tier infrastructure. |
Document Upload
│
▼
Classifier ─────── Groq · runs FIRST, its category feeds the rest
(gpt-oss-120b)
│
┌────┴──────┬──────────────┐ asyncio.gather()
▼ ▼ ▼
Safety Entity Summarizer ← skipped when the category
Analyzer Extractor doesn't warrant it
(Mistral) (Cerebras) (Cerebras)
└────┬──────┴──────────────┘
▼
Chunks + Embeddings → pgvector (HNSW)
Compliance Auditor (Groq) runs separately, ON DEMAND
→ per-clause Pass / Fail / Not Addressed matrix
Retrieval pipeline — query → 23 injection-guard patterns + 1,500-char cap → Gemini
gemini-embedding-001 (768-dim) → pgvector cosine (HNSW) fused with Postgres full-text
ts_rank_cd via Reciprocal Rank Fusion (k=60) → over-fetch 20 → ms-marco-MiniLM-L-6-v2
cross-encoder rerank to top 5 → streamed over SSE with inline [Document, Page X]
citations.
The lexical arm is PostgreSQL full-text search, not true BM25 — real BM25 needs an extension like
pg_search. Everyone writes "hybrid BM25 + vector"; I'd rather be precise than round up.
Retrieval quality — a blocking CI gate, scored on 12 labelled queries over a 130-chunk mining corpus. If it drops below the floor, the build fails:
| Metric | Floor | Current |
|---|---|---|
| Hit Rate@5 | 0.90 | 1.000 |
| MRR | 0.75 | 1.000 |
| Recall@5 | 0.85 | 0.958 |
| nDCG@5 | 0.75 | 0.968 |
242 tests run green as blocking gates on every push (215 unit + 27 integration); 274 collected once the eval suites are counted · 27.1K lines across two apps · $0/month infrastructure.
| Jul 2026 – Present |
Open Source Contributor · OmniRoute — Remote Universal AI gateway — 53k★, 230+ LLM providers, 21,000+ tests. 6 PRs merged across backend schema design, provider integrations, routing logic and frontend. |
| Oct – Dec 2025 |
Full Stack Developer Intern · nTheta Works Pvt. Ltd. — Remote Engineered a two-stage semantic retrieval pipeline (Ollama embeddings → Redis HNSW → FlashRank reranking) for NLPForge, an enterprise LLM API-testing platform — +40% template-matching accuracy, −60% manual QA effort. Shipped async FastAPI microservices and Next.js/TypeScript dashboards, containerized with Docker Compose and CI/CD on Linux. |
| Jul – Aug 2025 |
AI & Full Stack Developer · Freelance — Remote Built SmartLearnX, an AI-powered LMS with dropout prediction (Logistic Regression, 91.4%) and performance forecasting (Random Forest, R² = 0.89) deployed as a FastAPI microservice behind a React/Node.js frontend. Added BERT quiz generation and a spaCy chatbot serving sub-2-second responses under load. |
| May – Jul 2025 |
Full Stack Developer Intern · OBG Outsourcing Pvt. Ltd. — Jaipur Led FinSageAI360, a multi-tenant financial-intelligence SaaS — cut monthly close reporting time by 45% and manual operational effort by 30% via AI anomaly detection and real-time KPI dashboards. Designed a JWT-authenticated REST API (Node.js/Express/MongoDB) with granular RBAC for multi-tenant data isolation. |
| Jun – Aug 2024 |
Software Developer Intern · Om Logistics Ltd. — Delhi Optimized enterprise document search with LangChain + FAISS embeddings — −70% query latency across 10,000+ documents and +40% retrieval accuracy. Built Node.js REST APIs automating logistics workflows, eliminating 20% of manual data entry. |
A 53k★ universal AI gateway: one OpenAI-compatible endpoint over 230+ providers, MCP server, A2A protocol, 21,000+ tests.
| PR | What it took | Impact |
|---|---|---|
| #9397 | Enforce gemini-web reasoning and tool constraints — provider capabilities modelled declaratively rather than patched at the call site |
Fixes malformed requests to a strict provider |
| #9353 | Correct reset-window strategy prioritization in the router | Routing picks the right provider under quota pressure |
| #6366 | Dynamic import for MITM + fixed Turbopack over-bundling warnings | Smaller bundle, clean build output |
| #6245 | Accessible "Configured Only" filter for provider rankings — live /api/providers state → filterable grid, role="switch" + aria-checked, useEffect cleanup |
168 additions / 4 files, 9/9 tests, shipped v3.8.45 |
| #6209 | Claude 5 Sonnet into the claude_web provider registry with a regression test |
Signed commit merged within hours of the model's release |
| #6105 | Audited 9 core docs + 20+ localized READMEs across 42 locales; corrected stale architecture facts (routing strategies 13→17, service modules 36→134) | Passed docs-sync-strict, zero regressions, v3.8.44 |
The one I'm proudest of — #6225: diagnosed an HTTP 400
regression hitting strict LLM providers (Xiaomi MiMo), caused by system-message ordering in
the memory-injection pipeline. Rather than special-casing the provider, I proposed a
declarative Zod schema flag — systemMessageMustBeFirst — which the maintainer adopted
into the broader shipped fix. 25/25 Vitest + 30/30 Node test-runner coverage.
Languages
AI / ML
LLM Providers
Backend & Frontend
Data & Infra
The card above is generated from the GitHub API by
scripts/gen_stats.py and refreshed weekly — a
committed file rather than a live third-party widget, because the public
github-readme-stats instance is currently paused (HTTP 503) and its cards were
rendering as broken images here.
Long-form engineering write-ups on what actually broke and why:
- 📝 How I Built a Production-Grade Multi-Agent AI System That Won a National Hackathon — the 5-agent architecture, and what real-world GenAI cost me to learn
- 📝 Stop Overpaying for Vector DBs: A Production Hybrid RAG Pipeline for $0/Month — pgvector + Postgres full-text + RRF, and where the free tiers actually bite
- 📝 The State Machine Paradigm: Why I Ditched Linear LLM Chains for LangGraph — per-agent error isolation, conditional branching, fallback routed by token budget
- 📝 RSC + Streaming LLMs: A Zero-Latency AI Dashboard with Next.js Server Actions
B.Tech, Computer Science & Engineering — Global Institute of Technology, Jaipur Oct 2022 – Apr 2026 · CGPA 8.10/10
- 🏆 Smart India Hackathon 2023 — National Winner · Ministry of Coal, top 1% of 44,000+ teams
- 📄 Scopus-Indexed Publication — PiCET-2026 (IET Conference Proceedings): hybrid attention-based temporal modeling
- 📜 NASSCOM Certified Full Stack Developer (2024)
- 🎪 CodeFiesta Organizer — sponsorship lead across editions 3.0 and 4.0
