Builds

Things I've made.

From a pipelined CPU to trust infrastructure securing GPU workloads: the projects I've built, broken, and rebuilt.

March – May 2025

RISC-V CPU in Verilog & Logisim

A 3-stage pipelined RISC-V CPU, prototyped from scratch and verified like production hardware.

VerilogSystemVerilogLogisim
  • Designed a 3-stage pipelined RISC-V datapath (Fetch, Decode/Execute, Memory/Writeback) with a full register file and PC control logic, prototyped in Logisim.
  • Wrote SystemVerilog assertions and directed testbenches to formally check datapath, register file, and PC-update correctness cycle by cycle.
  • Used waveform-based debugging to isolate and fix faulty PC updates, misaligned instruction fetches, and memory-timing bugs across the pipeline.
View on GitHub ↗
Hand-drawn notes and diagram of the 3-stage pipelined RISC-V datapath, covering instruction fetch, decode/execute, and memory/writeback stages

Summer 2025 · Nutanix

GPU & Host Trust Attestation Infrastructure

End-to-end trust infrastructure securing confidential AI/ML workloads across physical hosts and GPUs.

TPMTEENVIDIA H100AMD SEV-SNP
  • Implemented TPM 2.0 remote attestation using PCR-based measured boot and cryptographic verification, enabling secure access control for AHV hosts and NVIDIA H100 GPUs.
  • Deployed end-to-end attestation infrastructure integrating cloud flows (AMD SEV-SNP, NVIDIA Trust Outpost) with per-session secrets locked behind AIK credential activation and PCR validation.
Nutanix and GPU trust attestation infrastructure project cover

Summer 2025 · Citylitics

18M+ URL Data Integration Pipeline

Automated extraction and curation pipeline that turned a noisy web crawl into a gold-standard ML dataset.

PythonSQLBigQuery
  • Built an automated extraction pipeline from an 18M+ URL web crawl into BigQuery, cutting noise by ~70% and halving query runtime.
  • Designed a human-in-the-loop labeling workflow with anomaly detection to catch mislabeled and out-of-distribution records.
  • Delivered a gold-standard curated dataset that accelerated downstream ML dataset preparation across the team.
18M+ URL Data Integration Pipeline project cover

Student project

Formula Electric Battery Management System

Battery management system work for a Formula Electric racing team.

Embedded CPCB Design
  • Designed KiCad schematics for pack-level battery temperature sensing, per-cell voltage sensing, and transistor-switched cell balancing.
  • Built voltage-sensing circuitry around LMV331 comparators feeding the MCU's ADC inputs for real-time cell monitoring.
  • Mapped the MCU pinout and embedded C interfacing between the sensing hardware and the accumulator EE stack for the team's electric race vehicle.
KiCad schematic sheet showing battery temperature sensing, cell balancing with transistor switches, and voltage sensing circuitry

Student project

Two-Stage Op-Amp

Analog circuit design and simulation of a two-stage operational amplifier.

Analog IC Design
  • Designed a two-stage op-amp: a telescopic cascode first stage for high gain and a class AB second stage for low quiescent current, linked by a Miller compensation cap.
  • Sized transistors and a 7.25:1 current-mirror bias to hit a closed-loop gain of 2 while meeting a 180ns settling-time spec.
  • Simulated 65.4° phase margin, 0.044% total error, and 138.3ns settling time at 801.6µW, beating the target static and dynamic error budgets.
View on GitHub ↗
Schematic of the two-stage op-amp: telescopic cascode first stage and class AB output stage with Miller compensation

Student project

SIXT33N: Voice-Controlled Car

A voice-controlled car built for an embedded systems design course.

Embedded SystemsSignal Processing
  • Characterized motor dynamics via a PWM duty-cycle sweep and encoder logging, fitting per-wheel linear velocity models (θ, β) to correct for wheel asymmetry.
  • Built closed-loop control on wheel-distance differential (δ[i] = d_L[i] − d_R[i]) for straight-line driving and controlled 90° turns, with startup jolts to overcome static friction.
  • Trained a PCA/SVD voice classifier (3 principal components, ~40 samples per word) to recognize 4 spoken commands and trigger the matching drive maneuver in real time.
View on GitHub ↗

Final demo: SIXT33N listening for a voice command and driving the corresponding maneuver.