I am an AI/ML Engineer and full-stack builder who designs, builds, and ships production machine learning and Generative AI systems end-to-end — from hand-writing tokenizers and training transformers from scratch, to fine-tuning production language models, to architecting RAG pipelines and tool-using agents with rigorous evaluation.
B.Tech CSE (AI/ML) student at SGT University (GPA 8.8/10). Co-founder & lead developer of ReVive, an AI physiotherapy platform incubated under IHFC, IIT Delhi. I pair research-grade ML with product-grade engineering — clean service-layer architecture, real-time systems, and deployed APIs people actually use.
- AI / ML — LLM pretraining from scratch, PEFT & DAPT fine-tuning, RAG, agentic RAG (LangGraph), tokenizer design
- Software Engineering — real-time systems (WebRTC, Socket.IO), booking engines, monorepos, CI-ready repos
- Full-Stack — Next.js · React Native · FastAPI · Node/Express · MongoDB · Redis · Supabase
- MLOps / LLMOps — FastAPI · Docker · Hugging Face Spaces · Vercel · RAGAS · DeepEval · Langfuse
Open To — AI Engineering internships · Applied research · Open-source collaboration
Languages
Frontend
Backend & Databases
Cloud, DevOps & Tooling
| Domain | Proficiency | Details |
|---|---|---|
| LLM Pretraining (from scratch) | ▰▰▰▰▰▰▰▰▰▱ |
5.19M-param decoder-only transformer · val perplexity 2.51 (Katha-LM) |
| Tokenizer Design (BPE) | ▰▰▰▰▰▰▰▰▰▱ |
Custom Hindi BPE · vocab 756 · 1.63 fertility vs GPT-2's 7.02 |
| LLM Fine-tuning & DAPT | ▰▰▰▰▰▰▰▰▰▱ |
xlm-roberta DAPT + 3-stage fine-tune · Macro F1 0.9304 (Artha) |
| Agentic RAG (LangGraph) | ▰▰▰▰▰▰▰▰▱▱ |
Self-correcting agent over 491K refs · 62.2% MIRAGE (Medical RAG) |
| RAG & Reranking | ▰▰▰▰▰▰▰▰▰▱ |
Cross-encoder rerank · embedding ablation (general vs MedCPT) |
| LLM Eval & Observability | ▰▰▰▰▰▰▰▱▱▱ |
RAGAS · DeepEval · Langfuse · resumable eval harness |
| Agentic Workflow Automation | ▰▰▰▰▰▰▰▰▱▱ |
n8n agents · structured-output schemas · confidence flagging |
| MLOps & Deployment | ▰▰▰▰▰▰▰▰▱▱ |
FastAPI · Docker · HF Spaces · Vercel · model versioning + auth |
Katha-LM — Hindi Story Language Model (from scratch)
A from-scratch Hindi storyteller: a hand-written byte-level BPE tokenizer and a 5.19M-parameter decoder-only transformer that writes children's stories. No pretrained anything — tokenizer, model, training loop, and deployment built end-to-end.
| Stack | Python · PyTorch · Custom BPE Tokenizer · FastAPI · Docker · Next.js |
| Scale | 5.19M params · 100k Hindi stories · trained on Kaggle T4 · tokenizer vocab 756 |
| Performance | Val perplexity 2.51 (bigram 26.8 → MLP 16.3 → transformer 13.1 → scaled 2.51) · fertility 1.63 vs GPT-2's 7.02 (~4× fewer tokens/word) |
| Security | Open weights · reproducible pipeline · private HF Hub model hosting |
| Impact | Shipped to production — FastAPI on HF Spaces (Docker) + Next.js 14 on Vercel |
| Repository | github.com/gourav-shokeen/katha-lm |
Demonstrates mastery of the full language-modelling stack — from byte-level tokenization design to transformer internals to production deployment.
Artha — Hinglish Sentiment Analysis API
A deployed production sentiment-analysis API for code-mixed Hinglish, powered by a domain-adapted and fine-tuned XLM-RoBERTa model, served behind FastAPI with a Next.js frontend.
| Stack | Python · PyTorch · xlm-roberta · HuggingFace · FastAPI · Docker · Next.js |
| Scale | 300k+ real Hinglish comments scraped & processed |
| Performance | 3-stage fine-tune: DAPT → v1 F1 0.7856 → neutral upsampling (lr 5e-6) → v2 Macro F1 0.9304 |
| Security | Bearer-token auth · model versioning · private HF Hub model hosting |
| Impact | Live production REST API at artha.social |
| Repository | artha.social · huggingface.co/gourav011/artha-v1 |
The full applied-ML lifecycle: data → domain adaptation → multi-stage fine-tuning → evaluation → deployed, authenticated API.
RFQ-to-Quote Agent — B2B Sales Automation
An n8n agent that turns inbound RFQs (PDF / Excel / email) into branded PDF quotes end-to-end, combining LLM extraction with deterministic catalog pricing and a human-in-the-loop approval gate.
| Stack | n8n · Gemini (Flash-Lite / Flash) · LlamaParse · Playwright · Telegram · Gmail · Google Sheets |
| Scale | 75-SKU catalog · multi-format ingestion (PDF / Excel / email) |
| Performance | LLM line-item extraction → in-context catalog matching & pricing → Playwright PDF render |
| Security | Telegram human-approval gate · confidence-based exception flagging · Google Sheets audit log |
| Impact | Inbound RFQs → branded quote PDFs + in-thread Gmail reply, fully end-to-end |
| Repository | github.com/gourav-shokeen/rfq-quote-agent |
Routes extraction/scoring to Flash-Lite and drafting to Flash, with strict structured-output schemas so low-confidence matches surface for review instead of failing silently.
Medical Reference Agent — Self-Correcting Agentic RAG
A self-correcting agentic RAG system (LangGraph state machine) over 491K clinical references that cites sources, refuses out-of-scope questions, and rewrites its own weak queries.
| Stack | Python · LangGraph · LangChain · Chroma · FastAPI · Next.js |
| Scale | 491K clinical references · 491K snippets embedded on a 6 GB GPU |
| Performance | 62.2% on MIRAGE medical-QA (7–8B model; GPT-3.5 ref. 71.6%) · MedCPT ablation surfaced a 4.4-pt retriever/reranker misalignment |
| Security | Cites sources · refuses out-of-scope queries · self-rewrites weak queries |
| Impact | Resumable RAGAS / DeepEval / Langfuse evaluation harness |
| Repository | github.com/gourav-shokeen/medical-rag-agent |
Built for high-stakes domains where every answer must be grounded, auditable, and aware of its own limits.
Invoice Validation Agent — LangGraph Self-Verifying Agent
A LangGraph agent that extracts invoice data with Gemini and validates that totals add up — retrying automatically, then flagging for human review when it cannot resolve discrepancies.
| Stack | LangGraph · Gemini 2.5 Flash · 5-node StateGraph |
| Scale | Self-verifying loop · passed 8 / 8 acceptance criteria |
| Performance | Structured extraction + arithmetic validation · auto-retry on failure |
| Security | Validation-gated output · escalates to human review on unresolved errors |
| Impact | Automated invoice QA with built-in correctness guarantees |
| Repository | github.com/gourav-shokeen/invoice-validation-agent |
A tight example of reliable, self-checking agent design built to a strict acceptance spec.
chaoslake — Synthetic Relational Data Generator
A CLI and Python library that generates relational synthetic datasets with realistic chaos — nulls, duplicates, drift, and messy formatting — from a simple YAML schema.
| Stack | Python · Polars · DuckDB · YAML schema · CLI + library |
| Scale | ~5.8M rows/sec · Polars (<500k rows) / DuckDB (>500k rows) backends |
| Performance | Configurable chaos: nulls, dupes, drift, format noise |
| Security | Deterministic seeds · reproducible synthetic data generation |
| Impact | Published on PyPI — pip install chaoslake |
| Repository | github.com/gourav-shokeen/chaoslake |
A polished open-source utility for testing data pipelines against realistically imperfect data.
Co-Founder & Lead Developer — ReVive 2026 — Present · IHFC, IIT Delhi
AI physiotherapy platform incubated under IHFC, IIT Delhi — a cross-platform patient + therapist application.
- Co-founded and lead development of a cross-platform app: 7-step clinical onboarding, real-time booking against live availability, and in-session WebRTC video consultations over a Socket.IO signalling namespace
- Engineered the booking engine with 30-minute slot generation and a 3-layer double-booking guard (Redis lock + DB uniqueness constraint + partial-index backstop)
React Native · Node.js · Express · MongoDB · Redis · Socket.IO · WebRTC
Event Coordinator — SGT University · Tuskers Club Feb 2026 — Apr 2026 · Gurugram
Led end-to-end planning and execution of campus debate and public-speaking events.
- Planned and executed 5+ public-speaking & debate events end-to-end; managed logistics and multi-channel outreach
Leadership · Event Management · Communication
| Recognition | Details |
|---|---|
| Hackathon Winner | AI/ML Hackathon & Coding Competition — SGT University (org. Samatrix Consulting), 2026 |
| ACIC SGT Funding | Secured funding for Click2Drive, 2025 |
| IIT Delhi Ideathon | Top 1000 — College Youth Ideathon, 2025 |
| Academic Excellence | GPA 8.8/10 — B.Tech CSE (AI/ML), SGT University |
-8B5CF6?style=for-the-badge&logo=python&logoColor=white){kind=icon}
gourav_shokeen:
role: "AI/ML Engineer · Generative AI · Full-Stack"
education: "B.Tech CSE (AI/ML), SGT University · 2024–2028"
base: "New Delhi, India"
learning:
- "Tool-using & multi-agent orchestration (LangGraph)"
- "LLM evaluation & observability (RAGAS, DeepEval, Langfuse)"
- "Karpathy — Neural Networks: Zero to Hero"
building:
- "Production GenAI: from-scratch transformers → RAG → agents"
- "Deployed, callable AI APIs (FastAPI · Docker · HF Spaces · Vercel)"
exploring:
- "Retrieval quality, reranking & grounded generation"
- "LLM cost & latency optimization"
open_to:
- "AI Engineering internships"
- "Applied research & open-source collaboration"
"From notebook to a deployed API people actually use — engineering intelligent systems from first principles to production."