Pearl Franz

Pearl Franz

I build data pipelines for government platforms by day and take apart neural recordings in my own time. Trained in neuroscience and biomedical engineering at Carnegie Mellon, now a software consultant and data engineer in Cambridge, MA.

Fig. 1 Simulated raster, 64 units at three depths. Scale bar 100 ms. Move your pointer over it to drive the units beneath.

About

Signals, pipelines, and the space between them.

I'm a software consultant and data engineer at Faraday Software, where I've worked remotely since June 2024. Most of my days go into moving public-sector systems onto modern platforms: ETL pipelines, SQL, AWS, and low-code application work in Appian.

Before that I studied how brains and bodies produce signals at Carnegie Mellon, first for a BS in Neuroscience and then an MS in Biomedical Engineering. My research touched computational neuroscience, biophotonics and bioinformatics.

I still care about that science, so most of my side projects sit where the two meet. They take messy biological data, build careful tooling around it, and report what worked and what didn't, including the negative results.

Data
Python, SQL, ETL pipelines, dbt, DuckDB, Postgres
Cloud
AWS Glue, S3, EC2, Bedrock
Machine learning
PyTorch, scikit-learn, NumPy, SciPy
Applications
Appian (Certified Associate Developer), FastAPI, Streamlit

Projects

Ten pieces of work, shown with their limits.

Everything here is public on GitHub. Each write-up lists its methods, limits and results, good or bad. The small drawings are generated, not data.

Neural data and theory

  • Pl. 1 / Allen Brain Observatory Neuropixels

    spikescale

    Asks where in time a population's stimulus information sits. It decodes natural-image identity from mouse visual cortex and thalamus at bin widths from 250 ms down to 5 ms, across 38 area samples in twelve mice. Cortex reads images best at 50 ms bins in 24 of 25 samples, and nothing is gained below 25 ms.

    It started as a test of a multiscale residual VQ tokenizer for spikes. The tokenizer didn't beat residual k-means by the margin I set as the bar, and the write-up says so.

    • Python
    • PyTorch
    • Neuropixels
    • Decoding
  • Pl. 2 / Latent dynamics, in progress

    sparsedyn

    How many neurons do you need, and how coarse can your bins be, before latent dynamics stop being recoverable? This project treats that as an observability problem with ground truth. So far the Fisher-information prediction matches an oracle smoother within 0.013 R² at fine bins, and a learned Poisson LDS reaches the oracle with 12 to 96 neurons. Nonlinear and real-data phases are next.

    • Python
    • State-space models
    • Simulation
  • Pl. 3 / Spike sorting from scratch

    spikesort

    The classic two-stage design in readable NumPy and SciPy: detect, cluster and build templates, then explain the raw trace with matching pursuit to recover overlapping spikes. Benchmarked on simulated tetrode recordings where the answer is known, it reaches 0.98 to 0.99 accuracy across noise levels.

    • NumPy
    • SciPy
    • Signal processing
  • Pl. 4 / EEG motor imagery, in progress

    eegllm

    Compares EEGNet, a transformer and a hybrid CNN-transformer on four-class motor imagery from the PhysioNet EEG dataset, then has an LLM write an interpretation of the predictions.

    • PyTorch
    • EEG
    • Claude API
  • Pl. 5 / Theory reproduction, simulated manifolds

    manifold-capacity

    A from-scratch reproduction of Chung, Lee and Sompolinsky (2018), Classification and Geometry of General Perceptual Manifolds, in NumPy, SciPy and cvxpy. Manifold capacity is how many object manifolds per neuron a linear readout can still separate. Simulated capacity came out at 1.92 to 2.03 for single points against Cover's 2. Spheres were 3.7% off the formula on average and 12% at worst. Point clouds were 3.2% off the estimator on average and 4.8% at worst.

    It also has three small extensions. A non-negative least squares version of the estimator gives the same answers as a generic cvxpy solve, 48 to 154 times faster per sample. With strongly correlated manifold centers the plain estimator overestimates (0.63 against 0.32 simulated), and removing the shared directions first gives 0.29. In a toy trained network with three hidden layers, estimated capacity rises from 0.23 at the input to 0.68 at the last hidden layer.

    It uses simulated manifolds, not neural recordings, at zero margin only and with 200 neurons at most. The network result comes from the estimator alone, and the network was tested on the same 30 objects it was trained on, so it checks the pipeline, not generalization. Applying it to Allen Brain Observatory Neuropixels data is the planned next step and hasn't been done.

    • NumPy
    • SciPy
    • cvxpy
    • Simulation

Research tooling

  • Pl. 6 / Image forensics

    dupscope

    Splits figures into panels and flags images that appear twice, even rotated, mirrored, cropped or re-toned, then writes a report for a person to check. On 200 held-out synthetic papers it found 58 of 80 planted reuses and wrongly flagged 1 of 99 clean papers.

    • OpenCV
    • SIFT
    • Benchmarking
  • Pl. 7 / Neuroimaging metadata

    neurolake

    Pulls BIDS datasets from OpenNeuro, lands the metadata as Parquet and models it with dbt on DuckDB. An audit of 150 random OpenNeuro datasets found that 51% pass every rule and 19% have no participants file.

    • dbt
    • DuckDB
    • Parquet
    • BIDS
  • Pl. 8 / Diffusion MRI, in progress

    dmriwebsite

    A SvelteKit site describing tractography, connectometry and longitudinal diffusion MRI analyses, with an inquiry form.

    • SvelteKit
    • TypeScript

Applied machine learning

  • Pl. 9 / Retrieval over PDFs

    ragpdfquery

    A production-shaped retrieval system with citations, multilingual support and hallucination checks. A LangGraph agent retrieves, grades, generates and retries with more context when an answer isn't grounded. Runs on Bedrock or OpenAI, with an admin dashboard for ingestion and evaluation runs.

    • LangGraph
    • pgvector
    • FastAPI
    • DeepEval
  • Pl. 10 / Vision in bad light

    lowlightlowresocr

    A fast, coarse pathway and a slow, detailed one, fused by a learned gate, for signs and text in bad light. On a task built so neither pathway can solve it alone, the fused model reaches 79.6% against about 51% for either pathway. On other benchmarks the gate mostly leaned on one pathway, which the README reports plainly.

    • PyTorch
    • GTSRB
    • OCR

Experience

Where I've worked.

  1. 2024 to now

    Software consultant and data engineer, Faraday Software

    Remote

    • State government client. Platform modernization project, current.
    • U.S. Department of Labor. Embedded on a federal project team, earlier engagement.

    Work across ETL pipelines, SQL, AWS (Glue, S3, EC2) and Appian application development.

  2. Before 2024

    Research, Carnegie Mellon University

    Projects in computational neuroscience, biophotonics and bioinformatics during my undergraduate and graduate degrees.

Education and certification

Contact

Say hello.

I'm happy to talk about data engineering, neuroscience tooling, or anything on this page.