# Sujal Maheshwari

> Full-Stack AI Developer building production-grade LLM systems, training language models from scratch, and shipping agentic infrastructure.

Portfolio of Sujal Maheshwari, a final-year B.Tech CS (AI & DS) student, currently Artificial Intelligence Intern at BeastLife (Gurugram, Haryana · On-site). The site is a single-page app: fetching its HTML returns an empty shell, so this file and https://sujalmaheshwari.com/llms-full.txt carry the content as plain text.

Contact: sujalmaheshwari07@gmail.com. Résumé: https://drive.google.com/file/d/10RppnxB6hD42rfWfY2UaftwrW6UB-Fq4/view?usp=drive_link

## Site

- [Home](https://sujalmaheshwari.com/): about, experience (https://sujalmaheshwari.com/#experience), selected work (https://sujalmaheshwari.com/#projects), stack (https://sujalmaheshwari.com/#stack), open source (https://sujalmaheshwari.com/#opensource), contact (https://sujalmaheshwari.com/#contact)
- [Writing](https://sujalmaheshwari.com/writing): index of technical notes and essays
- [RSS](https://sujalmaheshwari.com/rss.xml): the writing feed

## Case studies

- [Bot Street](https://sujalmaheshwari.com/work/bot-street): A full algorithmic-trading simulator on Kafka — price-time-priority order book, circuit breakers, 15+ quant indicators, and an MCP server so LLM agents can trade live against bots.
- [DriftGuard](https://sujalmaheshwari.com/work/driftguard): A semantic mistake-memory for AI agents — it remembers what went wrong and warns before the agent repeats it.
- [Librarian Series](https://sujalmaheshwari.com/work/librarian): A 125M-parameter language model built end to end — custom tokenizer, modern GPT, a resumable data pipeline, and a config-driven SFT framework.
- [Mafia A2A](https://sujalmaheshwari.com/work/mafia-a2a): A Mafia engine with a custom agent-to-agent protocol where choosing a private channel is itself a visible, suspicious move — plus a live WebSocket spectator server.
- [ToolStore](https://sujalmaheshwari.com/work/toolstore): One plain-English sentence in, one runnable MCP server out — via retrieval, reranking, an AST security scan, and codegen.

## Writing

- [A multi-agent framework where the graph builds itself](https://sujalmaheshwari.com/writing/a-multi-agent-dag-framework): Coven takes a task in plain English, decomposes it into specialized agents, infers the dependency graph from how they share artifacts, runs each level in parallel, and compiles one coherent result — wiring real tools per node along the way.
- [Predictive observability — drift-triggered retraining between Prometheus and Grafana](https://sujalmaheshwari.com/writing/predictive-observability-prometheus-grafana): Sentinel scrapes your metrics, trains a lightweight time-series model per metric, and emits forecasts back as standard Prometheus metrics — so Grafana plots real and predicted side by side, with zero changes to either. The interesting part is what happens when a model goes stale.
- [Training a 125M-parameter language model from scratch](https://sujalmaheshwari.com/writing/training-a-125m-language-model-from-scratch): What it actually takes to go from a folder of text to a coherent GPT — a modern architecture (RoPE, RMSNorm, SwiGLU) and a restart-safe data pipeline where losing six hours to a flaky disk is simply not allowed.
- [Designing an agent-to-agent protocol for social deduction](https://sujalmaheshwari.com/writing/designing-an-agent-to-agent-protocol): Agents that play Mafia have to talk like people do — broadcast, huddle, whisper — and a private channel should itself be a visible, suspicious move. Here's the A2A protocol I built for it, and the server that streams a fresh game every hour.
- [Turning plain English into a runnable MCP server](https://sujalmaheshwari.com/writing/building-an-mcp-autobuilder): ToolStorePy takes a one-line tool description and produces a single runnable MCP server — via semantic retrieval, cross-encoder reranking, an AST security scan, and codegen. Here's the pipeline, and the parts that actually matter.

## Projects

- [DriftGuard](https://github.com/sujal-maheshwari2004/DriftGuard): Agentic Infra · Memory Layer. Semantic mistake-memory system for AI agents. (PyPI: https://pypi.org/project/driftguard-ai/, docs: https://drift-guard-ai-docs.vercel.app/)
- [Sentinel](https://github.com/sujal-maheshwari2004/Sentinel): MLOps Infra · Observability. MLOps pipeline that sits transparently between Prometheus and Grafana. (PyPI: https://pypi.org/project/sentinel-mlops/, docs: https://sentinel-docs-jet.vercel.app/)
- [ToolStore](https://github.com/sujal-maheshwari2004): LLM Infra · MCP Builder. Automatic MCP server builder. (PyPI: https://pypi.org/project/toolstorepy/, docs: https://tool-store-py-docs.vercel.app/)
- [Librarian Series](https://github.com/sujal-maheshwari2004): Personal Research · LLM From Scratch. 125M param causal LM built end-to-end: custom BPE tokenizer, GPT with RoPE + RMSNorm + SwiGLU, trained on WikiText-103 + TinyStories (6.19 perplexity).
- [librarian-press](https://github.com/sujal-maheshwari2004/librarian-press): LLM Infra · Training Framework. Config-driven framework to pretrain and fine-tune GPT models from a single JSON file. (PyPI: https://pypi.org/project/librarian-press/, docs: https://librarian-press-docs.vercel.app/)
- [Mafia A2A](https://github.com/sujal-maheshwari2004/mafia-a2a): Multi-Agent · A2A Protocol. Mafia (Werewolf) engine with a custom agent-to-agent comm protocol — unicast / multicast / broadcast with dual-layer visibility, where choosing a private channel is itself a visible, suspicion-fuelling move. (live: https://mafia-a2a-frontend-909729949991.asia-south1.run.app/)
- [Bot Street](https://github.com/sujal-maheshwari2004): Simulation · HFT Environment. Full algorithmic trading simulator on Apache Kafka (KRaft). (live: https://bot-street-terminal-957519594640.us-central1.run.app/)
- [Coven](https://github.com/sujal-maheshwari2004/coven): Agentic Infra · Multi-Agent DAG. Multi-agent DAG pipeline framework. (PyPI: https://pypi.org/project/coven-ai/)
- [PeakPulse](https://github.com/sujal-maheshwari2004): Project · AI Product. Customer support intelligence for e-commerce.
- [NewsCheck](https://github.com/sujal-maheshwari2004): Project · Pipelines. YouTube → yt-dlp audio → Whisper transcription → GPT-4 summary in 5 structured bullets.

## Links

- [GitHub](https://github.com/sujal-maheshwari2004)
- [LinkedIn](https://linkedin.com/in/sujal-maheshwari)
- [Hugging Face](https://huggingface.co/MaheshwariSujal)
- [Résumé](https://drive.google.com/file/d/10RppnxB6hD42rfWfY2UaftwrW6UB-Fq4/view?usp=drive_link)
- [Email](mailto:sujalmaheshwari07@gmail.com)

## Optional

- [llms-full.txt](https://sujalmaheshwari.com/llms-full.txt): everything here plus experience, stack, and the full text of every case study and post