John Abrahams

johnnycabrahams@gmail.com | github | linkedin | resume

Featured projects

Kronos: A Type II Exovisor with WebAssembly
2025 to present. Rust. Co-authored paper. Live at kronoslabs.dev.
A multi-tenant runtime that combines WebAssembly's sandbox with exokernel-style capability gating and a custom wasi:thread-spawn dispatcher. ~5,000 lines of Rust; ~0.08 MiB per isolate (3x denser than Wasmtime); up to 6.7x speedup on dense matmul while preserving per-tenant trap containment. Since July 2026 it also backs kronoslabs.dev, a service where a build turn yields a capability-gated app shareable by link.

Learning After Deployment
Sep 2026 to present. LLMs. Context distillation, LoRA, Pythia.
What objective should update a deployed language model's weights from new experience? The bet is context distillation under a KL forgetting constraint. On Pythia-410M and 1B: distillation stores facts rather than strings and forgets about a third as much as raw-text fine-tuning at matched recall; adding the replay term brings forgetting to near zero at 0.97 fact accuracy. Synthetic facts so far.

DJ Reinforcement Learning System
Dec 2025. Embedded. GRU + RL.
A wireless ESP32-S2 sensor mesh (environmental, audio, thermal) that streams crowd-engagement metrics to a Raspberry Pi controller. A 2M-parameter GRU pretrained on ~3,000 DJ sets predicts the next song; an RL head adapts to live sensor feedback. Built with Chris Spletzer for UPenn ESE 519.

Other projects

Recurrent Reasoning Grafts: A Negative Result
Jul to Sep 2026. LLMs. TRM/HRM, LoRA, Vast.ai.
Does a TRM-style recurrent module grafted into a frozen LM beat parameter-matched LoRA on multi-hop composition? No: LoRA won every held-out probe by 0.4 to 0.85 absolute accuracy across three seeds. Written up with a 49-claim evidence ledger and from-scratch controls.

DJ Copilot
2026. Audio ML. MERT, contrastive next-track model.
42,397 tracks embedded with MERT-v1-330M over a 1.07M-event corpus of DJ sets. Track-level modeling lost to an adjacency-count baseline (4.6% vs 6.9% recall@10). Time-indexed features for what the DJ community was playing each month then lifted validation recall@10 from 9.8% to 13.5%, and cold-track recall from 3.1% to 7.8%; a month-shuffle control returns the base, so the effect is timing, not fame.

Ninth Wave: Software Engineer Intern
Summer 2026. Fintech. AWS CDK, living documentation, MCP gateway, Spring Boot.
Automated bank onboarding onto Compass's AWS platform with a CDK generator (one YAML file per bank), built the pipelines behind Ninth Wave's self-updating documentation, built Studio (an OpenAPI-to-MCP gateway for internal apps), and shipped eleven changes to the Spring Boot portal banks log into.

Home Compute Cluster
Aug 2026. K3s, Raspberry Pi, SOPS, Grafana.
A mixed-architecture K3s cluster (two Raspberry Pi 4s, an Ubuntu iMac control plane, a Lima VM on an Intel MacBook) with encrypted secrets, verified snapshot restores, and monitoring over Tailscale. Runs the unattended crawl and embedding jobs for the music projects.

Tracklist Engine
2026. Audio ML. Rust + Python, DETR, Demucs, MERT.
A provenance-first DJ-set alignment engine over a 41,492-set corpus. Combines DETR cue-point detection, Demucs source separation, and MERT music-transformer embeddings over a normalized SQLite schema.

Clinical Therapy Chatbot
May 2025 to Mar 2026. LLaMA, FastAPI.
A production therapy chatbot for New York Brain & Spine Surgery, blending rule-based clinical workflows with LLM-generated responses. Configurable multi-agent system spanning multiple mental-health conditions.

The Quant Game
2024. Co-founded. React + Python + AWS.
A full-stack options-trading sim with 61M financial data points on AWS. Users build and backtest strategies and watch them play out against historical market data.

Dropbox Document Search Engine
2024. Capstone. NLP + AWS.
A search engine for ~23,000 Dropbox documents, with sentence-level highlight extraction. Partitioned indexing + Dropbox API integration for real-time access. Lehigh capstone for an early-stage startup.

neurovariety: Open-Source ML Library
2026. Rust + Python. Algebraic geometry. MIT license.
An open-source Python/Rust library that models neural-network function spaces as algebraic varieties, computing dimension, degree, and singular loci to quantify expressivity and implicit bias.

Hangover Simulator
Interactive. ODE model.
BAC kinetics, acetaldehyde clearance, sleep disruption, and an RL discount factor for drunk-you's choices.

Education

University of Pennsylvania, School of Engineering and Applied Science, Expected December 2026
M.S.E. in Data Science. GPA 3.62/4.0.
Coursework: Theory of Machine Learning, Emerging Theory of Large Language Models, Advanced Operating Systems

Lehigh University, P.C. Rossin College of Engineering and Applied Science, May 2025
B.S. in Computer Science and Engineering. GPA 3.65/4.0 (major 3.75). Minor: Probability and Statistics.

Experience

Ninth Wave, Software Engineer Intern, AI Team. May 2026 – Aug 2026

• Built a generator that turns one YAML file per bank into that bank's AWS stacks. Onboarding had meant about ten hand edits per bank, and all three production banks had shipped with a routing defect.
• Built the pipelines behind self-updating documentation: a weekly Lambda that reads the AWS accounts and rewrites the docs with Bedrock, and a scanner that drafts updates when code changes. Tested but not yet deployed.
• Built Studio, a gateway that lets an AI agent call any internal app through one MCP endpoint, with one set of rules for who may do what. Rewrote it after security review found two ways a caller could act as someone else.
• Shipped eleven changes to the Spring Boot portal that banks log into: an activity timeline with channels and search, a browser for the files banks keep in S3, and documentation white-labeled per bank.

New York Brain & Spine Surgery, Software Engineer. May 2025 – Mar 2026

• Built a production therapy chatbot on LLaMA and FastAPI. Clinical rules control the flow of each conversation and the model writes the replies, so one system covers several conditions.
• Co-authored a study of outcomes for 236 spinal-fusion patients as a research associate there in 2023.

Research

Kronos: WebAssembly Runtime, UPenn (advisor: Prof. Sebastian Angel). Oct 2024 – Present

• Built a WebAssembly runtime in Rust that runs many tenants in one process. Each tenant costs about 0.08 MiB of memory, about a twentieth of a Docker container, and gets its own CPU and wall-clock budget.
• Wrote a work-stealing thread pool for guest threads. Matrix multiply runs up to 6.7x faster on four cores.
• Co-wrote a nine-page paper on the design with Ajax Li (unpublished).
• Turned the runtime into kronoslabs.dev, a live site where a plain-language request becomes a sandboxed app.

Learning After Deployment, independent research. Sep 2026 – Present

• Studying how a deployed language model should learn new facts without forgetting what it already knows. The method: train the model to reproduce what it would have said with the fact in its prompt.
• Found on Pythia-410M that this stores the fact itself rather than the sentence it came from, and forgets about a third as much as ordinary fine-tuning at the same recall. The result holds at 1B.
• Added a penalty for drift on unrelated text, which brought forgetting to near zero while the model still answered 97% of questions about the new facts. Synthetic facts only so far; unpublished.

Recurrent Reasoning Grafts in Frozen Language Models, independent research. Jul 2026 – Sep 2026

• Tested whether adding a small recurrent reasoning module inside a frozen language model beats plain LoRA fine-tuning on multi-hop questions. It does not: LoRA won every test by 0.4 to 0.85 in accuracy across three seeds.
• Ran the control that could have rescued the idea: trained from scratch, the same blocks score higher when run once than when run in a loop. Wrote it up as a draft paper, unpublished.

Projects

DJ Copilot, independent audio-ML project. Jul 2026 – Present

• Embedded 42,397 tracks with a music transformer (MERT) and trained a model to predict the next track in a DJ set from 1.07M examples. It lost to a simple count of what usually follows each track (4.6% vs 6.9% recall@10).
• Added features for what the DJ community was playing each month, which more than doubled recall@10 for tracks with no play history, from 3.1% to 7.8%. Shuffling the months erased the gain, so the effect is timing, not fame.

Writing

Kronos: A Type II Exovisor Architecture with WebAssembly (2025, with Ajax Li)
Recurrent Reasoning Grafts Are Dominated by LoRA: A Negative Result (2026, sole author, draft)
Predictive Outcome Modeling After ACDF (LDR-C ROI Implant) (2024, co-author, 236-patient cohort study)
Gradient Descent Analysis (2023, with Prof. Arielle Carr, Lehigh)
A Highlighting Algorithm for Document Search (2024, capstone writeup)

More on the writing page.