ML / AI Engineer · MS ECE @ Carnegie Mellon (AI/ML) · Graduating May 2026 Looking for full-time ML / AI Engineer roles starting May 2026.
I'm a software engineer with 6 years of production fintech experience (Java/Spring Boot microservices at Fidelity Investments and Principal Financial Group), now finishing my MS at CMU specializing in machine learning. I care about ML systems that actually run, not just notebooks that work once.
Currently:
- Graduate Research Assistant at the CMU Language Technologies Institute, working on agent safety evaluation infrastructure with Prof. Graham Neubig and Prof. Maarten Sap.
- Graduate Teaching Assistant for 11-785 Introduction to Deep Learning (~400 students).
| Area | What that means in practice |
|---|---|
| Agent safety & evaluation | OpenAgentSafety benchmark infrastructure, trajectory analysis, real-time safety monitoring with Gray Swan Cygnal API |
| LLM systems | RAG, multi-agent orchestration (LangGraph), tool use via MCP, local + cloud LLM serving |
| Applied deep learning | Transformer ASR, multimodal retrieval, metric learning, generative models |
| MLOps & cloud | Vertex AI, GCP/AWS, Terraform, Kubernetes, model serving at scale |
🛡️ OpenAgentSafety Infrastructure — Sole engineer migrating the OpenAgentSafety benchmark to the OpenHands SDK. 359 LLM safety tasks across 8 risk categories, per-task Docker isolation, redesigned bash-accessible NPC system. Published HuggingFace dataset for open-source LLM safety eval. Advised by Prof. Graham Neubig.
→ Open-source contributions: OpenHands/benchmarks
🏆 Scotty-Scheduler — Multi-agent RAG course advisor built in 24 hours. Won Best Use of MistralAI at the CMU AI Agents Hackathon 2025. Stack: LangGraph · LlamaIndex · Pinecone · Mistral-7B · Modal.
🚕 NYC Taxi Fare + RAG Travel Assistant — End-to-end MLOps on GCP. XGBoost fare predictor (3.6 RMSE) trained with Vertex AI HyperTune, served via App Engine + 5 GCP services, with a LangGraph multi-agent layer routing between fare prediction and a Vertex AI RAG Engine powered by Llama 3.1-70B.
🧠 EEG-Multimodal-Project — Custom 1D Vision Transformer aligning 122-channel EEG signals to CLIP's text space via LoRA + Knowledge Distillation. Recall@5 of 17.52% (~2× baseline); 95% parameter reduction via LoRA. Trained on the Pittsburgh Supercomputer Center cluster. Team project — CMU 11-785.
🎙️ Voice Grocery Agent — Real-time voice agent with hot-swappable LLM backends (GPT-4o ↔ Ollama) exposed via MCP. LiveKit + Whisper + Silero VAD, ~3-5s end-to-end latency over a 3M-product catalog (Open Food Facts).
🔬 MyTorch + Transformer ASR — NumPy deep learning library implemented from scratch (autograd, CTC loss, beam search) + an encoder-decoder Transformer ASR system in PyTorch with joint CTC+CE loss, achieving 8.78% CER.
Detailed write-ups for each project are on my portfolio (link below).
🎬 multimodal-recsys-agent — Production-shaped two-tower retrieval + FAISS + LightGBM ranker on MovieLens 25M. Building in public — design rationale and progress visible in commits.
Also in flight: a culturally-aware image-to-music generative model
Languages Python · Java · SQL ML / DL PyTorch · HuggingFace Transformers · PEFT/LoRA · scikit-learn · XGBoost · LightGBM · FAISS LLM & Agents LangGraph · LlamaIndex · MCP · Pinecone · Vertex AI RAG Engine · Ollama · OpenHands SDK Cloud / MLOps AWS · GCP (Vertex AI, App Engine) · Azure · Terraform · Docker · Kubernetes · Helm Tools Weights & Biases · Modal · GitHub Actions · FastAPI · Spring Boot
Portfolio · LinkedIn · mgulavan [at] andrew [dot] cmu [dot] edu
If you're hiring for ML / AI Engineer roles starting May 2026 — or building something interesting in agent safety, applied LLMs, or ML systems — I'd love to hear from you.
Outside of tech: carrom, badminton, and currently learning crochet.