Jayce Walzer
Two surfers walking out at Cannon Beach, Oregon

Jayce Walzer

I build worlds for AI agents to live, play and fail in, then measure what they learn: RL environments, world-model simulations and multi-agent games.

walzerjayce@gmail.comresume (pdf)githublinkedin
Surfers waiting in the lineup at Short Sands, Oregon

selected work

Every demo is a real run. Press "how it works" for the explainer.

DealRoom 2026

A synthetic deal environment for tool-using agents, graded by a deterministic verifier: Qwen3 8B closed 13 of 32 deals, Llama 3.1 8B 1 of 46. Then SFT vs SFT + RL on Llama.

read more

DealRoom is a synthetic deal environment for testing and training tool-using language models. An agent has to close a fictional transaction: gather evidence, get authority, draft the right amendment, collect signatures on that exact revision before the deadline, commit to proceed or cancel, and report what happened. It works through seven typed tools, and a plausible explanation, an informal agreement or an unsigned draft does not count.

A deterministic Python verifier scores the final state: authority, signatures, deadlines, buyer constraints and the structured final claims, with no LLM judge. Inspect AI runs the agent loop, and a local React research console shows model results, trajectories step by step, and grading.

On 48 generated scenarios, Qwen3 8B passed 13 of 32 scored tasks and Llama 3.1 8B passed 1 of 46. Budget limits stay unscored, so the two cover different subsets. I then trained Llama with supervised fine-tuning and with SFT followed by RL; each passed the same 2 of 64 held-out tasks, so RL added no measured gain.

Python, Inspect AI, Ollama, MLX LoRA, React, TypeScript

Kusama studio 2026

A frozen vision-language model directing a frozen image model, every edit saved with its lineage. Its scale edit touched 1.65% of the image and changed none of the 1,031,297 protected pixels.

read more

Kusama Studio is the second version of Kusama. Instead of a hand-built medium, a frozen vision-language model directs a frozen image model. Qwen3-VL-8B-Instruct reads a reference packet and the current image, then proposes one concrete change and its expected effect. FLUX.2-klein-base-4B generates or edits the image. Nothing is trained, and both run locally on a 36 GB M4 Max, one after the other to fit in memory.

Every proposal, mask, image, hash and cost is saved with its parent, so each work carries a lineage tree. Model choices and human judgments are stored apart, and only a person advances a work.

The clearest result is a scale experiment. Qwen directed shrinking a hiker to one-third size. FLUX repaired a patch of 1.65% of the image, none of the 1,031,297 protected pixels changed, and the measured height ratio was 0.3365. Asked whether the smaller figure changed the feeling of vastness, the human reviewer said yes. Qwen's first unguided edit was less clear: the reviewer wasn't sure it felt any more vast.

Python, PyTorch (MPS), Diffusers, Transformers, FastAPI, SQLite

Shared whiteboard 2026

A real-time multi-user whiteboard built as a timed take-home: transactional versioning so racing edits never tear a shape. 26 of 26 concurrency checks with 5 users, 3 runs in a row.

read more

I built this real-time collaborative whiteboard as a timed take-home for Mechanize. It is meant for design reviews where 2 to 5 people on a call edit the same canvas, often the same shapes at once. It has accounts, invites, four shape types, live named cursors and per-user undo and redo, and canvases persist.

The hard part is keeping everyone consistent. Every edit is one Postgres transaction. It bumps a per-canvas sequence number under a row lock, so edits are serialized and that number becomes the shape's version. Only after the commit does the server broadcast the new row. Each client keeps the highest version it has seen, so a late or out-of-order frame can never move anything backwards. A shape's position and size are replaced as one unit, so a drag and a resize that race end as one user's shape, never a mix of both. Deletes are tombstones, so a late drag can't bring a shape back. Undo just sends the inverse of my own edit through the same path.

Before submitting, I ran a scripted test with 5 users on one canvas, firing edits at the same time: 26 of 26 checks passed on 3 runs in a row.

React, TypeScript, Vite, FastAPI, WebSockets, PostgreSQL (asyncpg), Docker Compose

Rent roll reader 2024

The core of OM-Analyzer, where I was CTO: an OCR pipeline that turns scanned broker rent rolls into a searchable tenant comps database.

read more

As CTO of OM-Analyzer (May–Jul 2024) I built a commercial real estate platform in Python and Flask. Its core turned offering memorandums into a tenant comps database. Brokers' PDFs often hold the rent roll as a scanned image with no text to copy, laid out differently by every broker.

You drag a PDF into the web app. It goes to S3 through a presigned upload, then the server finds the rent-roll page. It renders that page at 300 dpi, sharpens it, boosts the contrast and runs Tesseract. A cleaning pass drops note and boilerplate lines and keeps the lines that carry a dollar amount. It splits those lines at known tenant names, with anchor tenants held to their own rent thresholds. Each row is rebuilt from the OCR fragments: "1" "850" becomes 1,850 square feet, the lease end is matched to a 3, 5 or 10-year term, and rent per square foot is checked against annual rent. Rows land in a per-user master Excel sheet, with the property's address checked against a US ZIP table.

In the demo, a scanned seven-tenant rent roll goes in and seven complete rows come out. The sample is synthetic, and it ran locally against mocked AWS.

Python, Flask, Tesseract OCR, PyMuPDF, openpyxl, pandas, AWS S3 and Cognito, Stripe, Docker

Injectable PrEP model 2026

A week-by-week microsimulation and cost-effectiveness model of injectable HIV PrEP for people experiencing homelessness. Its sharpest finding: outreach without the injection is the most cost-effective arm.

read more

This is a transmission microsimulation and cost-effectiveness model of long-acting injectable PrEP (lenacapavir, with cabotegravir as a comparator) delivered to adults experiencing homelessness. It compares five arms: usual care, clinic delivery, two low-barrier outreach arms, and a control arm that gets the same outreach with no injection. The model tracks people week by week: infection through mixing groups, the care cascade, delivery capacity, refusal and missed visits, then costs and QALYs. Paired random streams make two identical arms come out exactly identical.

An important caveat: most inputs are hypothetical placeholders, and 107 of 118 parameter rows have no source. It is a structure for a decision, not an answer to one. Within those assumptions, three findings stand out:

  • Outreach with no injection (arm E) is the most cost-effective arm, at $84,691 per QALY over 5 years.
  • The break-even dose price runs from −$443 to $20,478, depending on what an averted infection is allowed to be worth.
  • Elimination can't be bought. Saturating delivery takes incidence from 1.337 to 0.163 per 100 person-years, and 63% of the infections left come from outside the population.

Python, PyTorch, NumPy, pandas, matplotlib, pytest

reels-edits 2026

A dashboard that cuts images to the beat as a seamless loop. In the demo every cut landed within 16 ms of the beat.

read more

Reels Edits is a local dashboard I built for making short, beat-synced image edits with Claude Code. Claude pulls candidate images only from a whitelist of sites, I keep or reject every one in the dashboard, and the order I keep them in becomes the edit order. Claude never finalizes an edit with an image I haven't kept.

The renderer does the timing. It fits the track's tempo and first-beat offset with librosa, sizes the edit to a whole number of bars between 7 and 12 seconds so the music keeps its rhythm when the video loops, and cuts on a repeating pattern of sixteenth notes. Every image gets the same grade, and the last shot is the kept image closest in colour to the first, so the loop seam is hard to see. A checker then measures the finished video against the rules.

In the demo, with 12 public-domain paintings and a generated 124 BPM beat, it found 123.96 BPM and made 12 cuts in 6 bars over 11.63 seconds. The worst cut was 16 ms off the beat, and every check passed.

Python, librosa, NumPy, Pillow, ffmpeg, Claude Code

A wave at Steamer Lane, Santa Cruz

experience

Fog clearing over Ocean Beach, San Francisco

education

honors

  • CreateX Launch
  • GT PACESHIP grant, $30K (about 10,000 GPU-hours)
  • ATDC compute grant
  • Top 250 in ultrabullet chess
  • Junior Olympian
  • Water polo, Chess Club

tools

  • PyTorch, RL, MCTS, RL environments and evals
  • quality-diversity search, LoRA, CLIP, LLM and VLM pipelines
  • Python, C++, TypeScript, React
  • PostgreSQL, AWS, Docker, Slurm and HPC
  • Claude Code, Codex, Cursor