Engineering
Engineering, Systems & Datasets
My research is supported by systems I build: agent runtimes, model connectors, experiment orchestration, logging, evaluation, deployment infrastructure, a published clinical video dataset, and lower-level hardware and verification projects.
01
AI systems and research infrastructure
The common substrate under the FORGE, compound-agent, and multi-turn studies.
Multi-provider LLM inference
Unified connectors across API providers with retries, pooling, rate-limit handling, and configuration management.
Agent harness and tool execution
ReAct-style hierarchical harnesses with bounded specialist interfaces and tool execution over simulator APIs.
Experiment orchestration
Parallel, seeded, resumable runs: 3,475 episodes in one study and 84,540 trajectories in another.
Configuration-driven agent definitions
Every condition, topology, and prompt variant declared in versioned configuration so a run is reproducible from one file.
Logging and token accounting
Structured per-step trajectory logs and per-call token accounting behind every cost figure reported.
Evaluation pipelines
Survival and hazard estimation, failure-rationale labeling with human audit, and cost-performance frontier analysis.
Reproducibility tooling
Archived Zenodo artifacts and documented reproduction paths. All ACM reproducibility badges on both CAIS papers.
Docker and GPU workflows
Containerized simulators and analysis. GPU training for sparse autoencoders and video transformers.
02
Applied ML systems and datasets
Infant Care Video Dataset
2024–2025A privacy-compliant NICU intervention dataset and a video-transformer benchmark.
- 4,144 videos across 12 classes
- TimeSformer and MotionFormer fine-tuned and benchmarked
- Up to 94% top-1 accuracy on neonatal interventions
Bandit-Driven Essay Grader
2025A multi-armed bandit agent that selects among grading prompts per essay to minimize inference cost.
- Four grading recipes with adaptive prompt selection
- Cost and reliability accounting per decision
- 78.4% reduction in LLM calls
Federated Anomaly Detection
2024A federated autoencoder pipeline for anomaly detection from network traffic logs.
- Containerized Mininet traffic simulation
- Federated autoencoder training across clients
- Network log analysis pipeline
03
Hardware and low-level systems
Secondary evidence of systems depth: design verification, CPU design, and a fabricated CMOS circuit.
RISC-V design-verification testbench
2025UVM-inspired scripted and classic class-based UVM self-checking testbenches with riscv-dv constrained-random stimulus.
- 8-seed regression with zero failures
- 88% line, 79% branch, 65% automatic Questa coverage (UCDB merged to an HTML dashboard)
- Two adder properties formally verified in SymbiYosys (depth 20)
RV32I single-cycle CPU
2022A synthesizable Patterson–Hennessy-style RV32I core in Verilog written from the official RISC-V specification.
- Xilinx simulation testbench with a 20-instruction assembly regression
- Covers arithmetic, branch, load/store, and jump paths with GTKWave traces
CMOS pseudo-random sequence generator
2022Full-chip layout of a 5-bit LFSR PRSG with an on-chip two-phase clock on the CUSoI SOI CMOS multi-project wafer.
- DRC/LVS closed in the Cadence design suite
- Fabricated and validated: 31-state ≈2.7 kHz output on a probe station
Contact
Building reliable AI systems requires both research and engineering.
I’m interested in research engineering, applied research, agent infrastructure, evaluation, reliability, interpretability, and research-to-production work.