PAUL IACOBUCCI

Research Assistant: SP26 - present. Profiling Mixture-of-Experts inference on 8×H100 GPUs for an industry client designing custom AI silicon.

Read the full writeup →

Software / ML Engineering Intern: Three internships across software, embedded, and FPGA. Most recently, built a real-time on-device neural TTS system for tactical radios (Piper + Sherpa-ONNX + ONNX Runtime, INT8-quantized) with an MQTT/LTE relay for out-of-range delivery — plus a training/eval pipeline that rejected a faster candidate model for failing an intelligibility check.

Software Engineering Intern: Full-stack development on a DoED-backed EdTech platform with a 3-engineer Cornell team.

A wearable hearing-protection app + firmware stack (nRF5340 + ADAU1860 DSP), evolved from a validated Teensy prototype. BLE app with live frequency tuning and a safety-capped hearing test; hardware bring-up in progress. View Site  ·  View GitHub

A from-scratch C++ inference compiler: parses ONNX protobufs into a custom IR, fuses Conv-ReLU to cut DRAM round-trips (11.2% on cache-exceeding inputs), and runs a trained MNIST CNN through handwritten NCHW kernels that match ONNX Runtime's logits exactly.

Try it live (WASM)  ·  View GitHub

A platform helping Cornell student-athletes network with and track relationships across 45,000+ Cornell athletic alumni. Solo full-stack build (Next.js + Expo/React Native) on a multi-tenant Postgres backend, shipped to 200+ active users.

View Site

Trained a CNN (PyTorch) on librosa features and ElevenLabs deepfakes, deployed via ONNX to mobile, Chrome extension, and desktop. Real-time inference with user alerts, shipped to 20+ users.

View Site

An FPGA trading pipeline that runs end-to-end in 444ns at 83.3M messages/sec. Register-partitioned limit order book, sparse MoE router pipelined to one trade per cycle, bit-exact RTL/C++ verification in Verilator.

View GitHub

Fused Triton kernels on H100. LayerNorm runs 45.7% faster with symmetric FP8 quantization. Scaled FlashAttention to 16K context by tiling for SRAM and computing softmax online in one pass.

View GitHub

An interrupt-driven tilt sensor and impact monitor on the FRDM-KL46Z (Cortex-M0+). Sleeps in __WFI between PIT timer wakeups; ARM assembly for the trig math; I2C accelerometer reads and UART alerting on impact.

Watch Demo Video

An interactive theorem prover in OCaml: logical proposition simplification, automated SAT solving, and a parsing/evaluation pipeline for mathematical proofs with automated LaTeX export.

Watch Demo Video

Built an AI-orchestrated modular chess engine evaluation system. Used Claude via an MCP server to autonomously test, benchmark, and compare diverse AI-generated chess engines using SPRT and perft.

View GitHub

Vocera: Biometric authentication and synthetic voice detection system built leveraging FastAPI, SpeechBrain, and OpenAI Whisper.

View GitHub

LockedIn: Professional networking application. Awarded Best UI.

View GitHub