Aaron G.

Most AI projects die in the gap between a demo that impresses and a system that survives contact with real users. I've spent 15 years on the production side of that gap. Recent builds: a fraud-detection pipeline with automated multi-model training, hyperparameter optimization, and ensembling, delivered ahead of a regulatory deadline at 79.4% accuracy. A multi-tenant document intelligence platform on React, FastAPI, and Qdrant doing sub-second semantic search across fully isolated client vaults. ML on line-level IoT telemetry that predicts injection molding defects in-cycle and attributes each call to the parameter drift causing it. A LangGraph agentic workflow that cut a screening process from 15 hours a week to 30 minutes. Full write-ups with architecture details are in my portfolio below. WHAT A CLIENT SAYS Tegra Design ran their clients' webinar programs by hand across Zoom, Gmail, and spreadsheets — staff pasting lead lists into email, watching for replies, registering attendees one at a time. I replaced three systems with one platform: upload a list, the campaign runs itself, registrations and reminders happen automatically. "Now we upload the list and the campaign runs itself. It gave my team their week back." — Nazr Irfan, CEO & Founder, Tegra Design WHAT I BUILD • Agentic LLM systems — tool use, multi-step reasoning, orchestration that doesn't fall apart on the third hop • RAG that's actually grounded — cited answers, role-based access, declines instead of guessing • Prototype to production — backend APIs, cloud deployment, evals, monitoring, the reliability work nobody demos • Applied ML — forecasting, predictive quality, NLP over messy operational text • Reinforcement learning for adaptive systems — agent training, dynamic model selection PEER-REVIEWED RESEARCH I pretrain and instrument transformers up to 1.5B parameters and publish on their internals. All three papers are on arXiv if you want to read them. "Three Roles, One Model: Role Orchestration at Inference Time to Close the Performance Gap Between Small and Large Agents" (first author) Running Qwen3-8B on a single 24GB GPU, we roughly doubled agent task completion on the AppWorld benchmark using inference-time scaffolding alone — no retraining, no extra compute. The scaffolded 8B model outperformed DeepSeek-Coder 33B Instruct, a system four times its size. Directly applicable if you're weighing frontier API costs against self-hosting. "Hidden Dynamics of Massive Activations in Transformer Training" (accepted, Nature Machine Intelligence) The first comprehensive analysis of how massive activations emerge across transformer training. We showed the emergence follows predictable mathematical patterns and built an ML framework that predicts them from architectural specs alone — meaning training instabilities can be anticipated before training starts. "When Model Merging Rivals Joint Multi-Task Reinforcement Learning: A Task-Vector Geometry Analysis" (sole author) Tested whether merging separately trained RL agents actually substitutes for joint multi-task training, a comparison the literature had skipped. It does — and the task-vector geometry explains why the choice of merging method barely matters. That research background isn't decoration. It's why I can tell you when the answer is fine-tuning, when it's retrieval, when a smaller model with better scaffolding beats a bigger one, and when your problem doesn't need an LLM at all — before you've spent the budget finding out. WHO I'M NOT FOR If you need a thin GPT wrapper or the lowest hourly rate on the platform, I'm the wrong call. Tell me what breaks today and what it costs you. I'll tell you honestly whether AI is the fix.