Eric Gilerson

Palantir Fellowship. I lead a team of 15 building an AI workspace for real estate development.

• Built Structora, a Tauri, React, and FastAPI desktop app on AWS (EC2, S3, MongoDB) for site files, deliverables, and team messages in one workspace instead of email, shared drives, and a separate chat.
• Cut first-draft time for briefs, reports, and RFIs by about 75% with parallel research workers over firm knowledge, the project, and the web, orchestrated on Celery and Redis, plus LanceDB hybrid search over prior drafts, firm files, and research.
• Built a plain-language question layer over 3D building models (IFC/BIM), raising useful answers from 50% to over 90%.
• Cached project context in Redis so tool-call loops reuse it, cutting assembly from 78ms to 9ms (9×) and off the database CPU.

Building compliance systems for enterprise AI agents.

1. Software Engineer, Part-Time Now

   Building core systems and client-facing applications for verified AI.

2. Software Engineering Intern Summer 2026

   • Shipped a Python SDK that finds running agents (over 99% accuracy) or registers them by hand, ingesting traces in FastAPI for runs, compliance evaluations, workflow graphs, version control, and identity management for every agent and workflow.
   • Cut the dashboard from 6.7s to under 100ms and each payload from 82 MB to 160 KB with Postgres indexes and JSON dedup.
   • Replaced a 5-second notification poll with event pushes, cutting p99 latency from over 15s to about 20ms.
   • Caught over 60% more risk and drift by filling incomplete agent cards with AI, and added failure explanations on all 42 rules.

Concert operations and ticketing for agencies and fraternity chapters. Events, payments, and scanning at the door, on Next.js, FastAPI, and AWS.

The short-video feed and live voice rooms for the Vex iOS app, mixing and watermarking more than 10 hours of video a day.

• Sole author of Don’t Learn What You Can Compute, accepted to NeurIPS 2026. Frozen residual blocks recognize a written equation and compute its exact integer result inside a transformer’s forward pass. The model learns only a small interface on top of a frozen base model, and the arithmetic itself stays fixed.
• On a frozen SmolLM2-360M, exact answers went from 4.56% to 99.95% on 36,000 supplied problems. Shrinking that added interface to about 306K parameters still scored 99.89% on the same problems, with the base model left frozen.
• Offloading the math reduced interference with a second task. Across five seeds, two equal models had to keep doing arithmetic while learning 40,000 unrelated facts. The one with exact arithmetic recalled those facts faster, peaking 7.5 points ahead at 10k steps, and its math stayed intact while the baseline’s slipped. Trained separately from scratch on text and equations, that same block raised math accuracy from 19% to 97.7% at the same text quality.

Current draft · Code

A Go and React desktop app in place of Docker Compose: one-click deploys, stable hostnames, services that point at each other by name, and multiple environments. An MCP server lets a coding agent fully manage the local environment. GitHub

A C++20 library on PyPI that turns per-ticker price CSVs into sliding-window training tensors and hands them to Python as NumPy arrays. PyPI · GitHub

A Fabric mod for Minecraft. Fly while riding a boat and set its speed, with a server config so an admin can lock flight or cap speed. About 54,000 downloads on CurseForge. CurseForge · GitHub

Pose estimation and a neural network that tell a good deadlift from a bad one, then turn the joint deviations into feedback a lifter can use. With Nikhil Lazarro, Ishaan Puri, and Razin Farooqi. GitHub · Preprint