I'm an AI systems engineer interested in the full stack of machine learning — from the systems and infrastructure that run models at scale to the research and training that make them better. I love working at the boundary where algorithms meet hardware constraints, where a smarter scheduler, a leaner architecture, or better fine-tuning can make all the difference.
I'm currently a Computer Science Honors student at Stony Brook University (GPA 3.9/4.0, Class of 2027) with hands-on experience building agentic AI infrastructure, LLM inference pipelines, and reliable concurrent systems.
- Bank of Montreal — AI & ML Engineering Intern (Jun–Aug 2026). As the sole engineer, built a 39-tool Python/FastAPI Model Context Protocol (MCP) server for an AML case-investigation pipeline from a blank scaffold in ~6 weeks — cutting a 3–4 hour manual workflow to 1–2 minutes (~99% faster), with ~$2M in projected savings.
- Reliable Systems Lab (Stony Brook) — Research Assistant (Jan 2026–Present). Researching multi-agent cyber-physical systems and neural controllers for UAV-swarm coordination; improved target-seeking efficiency by 34.6% using a CMA-ES-driven curriculum, and enabled 100% collision-free horizontal scaling on the SeaWulf HPC cluster.
- LUNR AI Lab (Stony Brook) — Research Assistant (Feb 2025–Mar 2026). Fine-tuned CodeLlama-7B via LoRA (460K+ custom samples), improving Coding RAG accuracy by 5.4% and accelerating benchmark runtimes by 73.5% with a parallelized vLLM routing system.
- Mailgator — Software Developer Intern (Sep 2025–Jan 2026). Architected a CI/CD pipeline that reduced validation latency by 99.9% (60 min → <2s) and refactored LLM parsing for 100% accuracy on edge cases.
- iGEM at Stony Brook (2024). Software engineer on a bio-computation team awarded a Gold Medal at the iGEM Paris 2024 Jamboree, building a RAG Q&A chatbot and Flask research wiki that secured $50K+ in funding.
I've published several open-source extensions for the Pi coding agent on npm:
| Package | Description | Downloads |
|---|---|---|
| pi-mtplx · npm | Run local MLX models with Pi — zero-config model discovery, server lifecycle, SSD session cache, live tokens/sec monitoring in the footer. | 1,800+/mo |
| pi-zg · npm | Native zvec-grep semantic code search for Pi agents, with always-fresh project context and agent-callable search tools. | 500+/mo |
- Research at the Reliable Systems Lab (Stony Brook) — multi-agent cyber-physical systems, neural controllers, and MPC for autonomous UAV swarms under adversarial conditions, with large-scale SeaWulf HPC simulations.
- Always exploring the agentic AI / MCP ecosystem and LLM inference infrastructure.
- Agentic AI: Model Context Protocol (MCP), agent/tool calling, RAG, LLM fine-tuning (LoRA/QLoRA), model distillation, vLLM, reinforcement learning
- Backend & Protocols: FastAPI, Flask, Node.js, React, REST/JSON-RPC, WebSockets, Server-Sent Events, SQLAlchemy
- Data & Platform: PostgreSQL, MongoDB, SQLite, DuckDB, Apache Parquet, ETL pipelines, AWS (EC2/S3), Docker, Linux/Unix
- Concurrency & Systems: async programming, multithreaded job scheduling, back-pressure, IPC, low-level systems, HPC (Slurm)
- Security & Automation: RBAC, JWT/API-key auth, OAuth/OIDC, hashing, audit logging, Playwright automation
Here are some of the projects I'm most proud of. You'll find more on my portfolio website!
| Project Name | Description | Tech |
|---|---|---|
| pi-mtplx | Local-first LLM runtime integration for Pi — automatic MLX model discovery, health-checked server lifecycle, thermal fan controls, live tokens/sec monitoring. | Python, Pi SDK |
| pi-zg | Native semantic code search for Pi agents via zvec-grep, respecting project search rules with always-fresh context. | Python, Pi SDK |
| REPLUG_LSR_with_vLLM | Refactored REPLUG for LM-Supervised Retrieval fine-tuning of code-generation models, driven by a local vLLM server. | Python, vLLM, PyTorch |
| Automated Systems Fuzzer | High-performance C fuzzer using Unix signals/syscalls (fork/waitpid) for 100% process isolation and automated memory-leak detection. | C, Unix |
| CMDFlow | Local-first AI command tracker (built at HackPrinceton) — streams shell activity with PII scrubbing and semantic search. | FastAPI, React, MongoDB |
| WindowsCenterStage | Emulates Apple's Center Stage on Windows with a custom-trained facial-recognition model that keeps your face centered. | Python, TensorFlow, OpenCV |
| C-Based Filesystem Emulator | In-memory Linux-like filesystem managing i-nodes and data blocks from scratch. | C |
| Multi-Client Poker Server | Concurrent server for real-time multi-client Texas Hold 'em gameplay. | Java, Networking |
| Red-Black Tree in MIPS Assembly | From-scratch Red-Black Tree in MIPS, demonstrating low-level memory management. | MIPS Assembly |
Stony Brook University — B.S. in Computer Science (Honors), GPA 3.9/4.0 (Aug 2023 – May 2027) Honors: SUNY SOAR Fellow, URECA Fellow, University Scholars, Dean's List (2023–2025) · Awards: Academic Achievement Award, YouAreWelcomeHere Award, Global Excellence Award
- 📧 Email: anshumaanvsingh@gmail.com
- 💼 LinkedIn: Anshumaan Singh
- 🏆 Devpost: KrossKinetic
- 📷 Instagram: @krosskinetic (Check out my photography!)
- 🧬 Google Scholar: Citations
- 🤗 Hugging Face: KrossKinetic
Inspired by Star Wars, my ultimate goal is to build a real-life C-3PO! 🤖