a-Fig
Bay Area, California

Tyler Darisme

Founder of HeyVi
Computer Science @ SJSU · Graduating May 2028
I like building cool things, hopefully you like them too. Lately that means developer tooling for AI coding agents. Accordion, my context management tool, has 246 stars on GitHub and won 2nd for best token compression at UC Berkeley's AI Hackathon 2026.
My Resume My GitHub My LinkedIn

Or email me directly at [email protected]

Animated fig character dancing

Highlights

Awards & Contests

2nd — UC Berkeley AI Hackathon

Jun 2026 · sponsor track, best token compression (Accordion)

1st — SJSU Programming Contest

Apr 2025 · Admitted Spartan Programming Contest

Top 20 — ICPC Regional

Feb & Nov 2024 · on-site

3rd — DAHACKS 3.5

May 2025 · AceAnything

Experience

Founder — HeyVi
Aug 2026 – Present

Building Vi, a personal secretary who lives on her own Android phone and runs on the customer's own Claude or ChatGPT plan (Kotlin, Node, Claude Code).

Consulting Project Lead — ACM @ SJSU
Sep 2026 – Present

Leading a client consulting engagement, coordinating a small team from requirements through delivery.

Tech Officer — Competitive Programming Club
Sep 2025 – Feb 2026

Delivered weekly algorithms & data-structures lectures; mentored members on contest prep.

SWE Intern — Shipcaptain
Jun 2025 – Aug 2025

Built a secure full-stack admin portal (Next.js 15, Prisma, PostgreSQL) with role-based access, audit logging, and NextAuth.

Blockchain Intern — Serraform Labs
Oct 2022 – Jul 2023

Shipped the on-chain USDC payment contract for a live Web3 product on a 7-person team.

Projects

HeyVi · My startup

Handled, by Vi.

I'm building HeyVi. Vi is a personal secretary who comes with her own phone, email and computer, and runs on the AI plan you already pay for.

  • Talks for you. She texts and emails people as herself, or as you.
  • Your AI, not ours. Her brain is your own Claude or ChatGPT plan.
  • Asks first. She never sends, pays or signs as you without your OK.
Join the waitlist heyvi.afig.dev

Investors and mentors: I'd love to talk, at [email protected]. Landing page built with @tthy-working.

Accordion — a visual map of an AI agent's context window

Accordion

246 TypeScriptTypes the folding engine, the Svelte desktop UI, and the pi extension that feeds it. TauriWraps the SvelteKit UI in a native desktop shell, so the context map is a real app rather than a browser tab. pi-coding-agentHooks pi's extension API to rewrite the outgoing messages before every model call, so the GUI can fold a live session.

Accordion is a developer tool for context visualization and intelligent context management that outperforms /compact. Where /compact erases history in one irreversible pass, Accordion folds individual blocks to cached summaries you can unfold instantly, and lets you pin the ones that must survive. An automated conductor handles the folding decisions by default, but you're always free to override it. My Thermocline conductor scored 83.3% vs 33.3% for naive compaction on SlopCodeBench. Won 2nd for best token compression at UC Berkeley's AI Hackathon 2026. Brand and design by @tthy-working.

GitHub Live

Automated Game Bot

27 $840K VolumeReal-money equivalent of the in-game silver a closed-source market-making offshoot of this bot has processed. Actively profitable. Pre-AIBuilt in 2024, before AI coding assistants were a major part of my workflow. PythonThe whole bot: the detection loop, the game control, and the threads that run them at once. PyAutoGUIDrives the real mouse and keyboard, so the bot plays through the game client the way a person would. Computer VisionFive YOLOv11 models read the screen live — resource nodes, mobs, obstacles, spell state, and health. RoboflowLabeled and trained all five detection models there, and published every dataset publicly.

A fully autonomous Albion Online bot written in Python. Five YOLOv11 models read the screen while it drives the real mouse and keyboard to farm, fight, and avoid danger. Scans and actions happen in parallel on separate threads so it always works from the most up-to-date information.

A closed-source offshoot applies a similar approach to market making. It has processed the equivalent of $840K in volume and is actively profitable, which is why the code is private.

GitHub Watch

Are You Smarter Than Claude?

2hr buildDesigned, built, and deployed in about two hours at Anthropic's Claude Impact Lab. TypeScriptTypes the Next.js app and the rules for all six games you play against Claude. Google Cloud RunDeployed as a container, so there is no signup — open the link and play.

A small project I built at Anthropic's Claude Impact Lab in Palo Alto, meant to be a fun way to make AIs like Claude feel more real and less scary. You can play six different games against Claude, including Connect Four, Battleship, trivia, and Wordle Race, and see firsthand where it is surprisingly good and where it completely falls apart. The goal is to give people a more intuitive feel for Claude's strengths and weaknesses. Built with Next.js and TypeScript, and hosted on Google Cloud Run. No account needed to try it, just click the link and play.

GitHub Live

AfterCart

TypeScriptTypes the whole path from receipt photo through to the store comparison. Next.jsNext 16 App Router: fifteen route handlers run the OCR and comparison pipeline server-side. Google Cloud RunRuns as a standalone container, talking to Cloud SQL over a Unix socket. Transformers.jsRuns MiniLM embeddings in-process to match receipt lines to products, instead of calling an embeddings API.

I built AfterCart at Davis Hacks 2026: photograph a grocery receipt and it OCRs every line item, then tells you what the same basket costs at nearby stores — matching against a crowdsourced database of 18,649 price observations across 244 stores, on Postgres 16 with PostGIS + pgvector, Google Vision OCR, and Gemini 2.5 Flash.

GitHub Live
Roadies on three phones: the jam screen with Connect and Random, a live voice room at SFO, and the same room in dark mode with drivers muted

Roadies

6hr buildBuilt in about six hours for a hackathon, from the design doc to a live deploy on Google Cloud Run. TypeScriptEnd to end: the React phone app, the Node server, and the shared package with the matchmaker, the command parser, and the protocol. LiveKitCarries the voice over WebRTC. Muting is a track-subscription permission rather than a muted mic, so the server can still hear you. Speech-to-TextGoogle Speech-to-Text hears each driver, and only an utterance that is exactly a command counts, so “don't mute me” does nothing. Works in English, French, Spanish, and Vietnamese.

Hands-free proximity chat for drivers stuck in traffic. You become a car, like “Teal Civic”, in a Discord-style voice room with the three drivers nearest you on the road, and you control it by voice: say “mute” and it mutes. Brand and design by @tthy-working.

GitHub Live

Jev Score

0 dependenciesThe package declares no runtime dependencies. The database, both servers, and the worker that runs custom scorers all come from Node's standard library. JevTypeSafe's scoring model. On TypeSafe's workflow evals it averages 0.4 s and $0.0004 a case, against 12.5 s and $0.0195 for Claude Haiku 4.5, and it is more accurate too (67.8% vs 53.6%). JavaScriptPlain ES modules and no framework, from the CLI and both servers to the web app. MCPServes the same features as Model Context Protocol tools, so chat apps without a shell, like Claude Desktop, can run the loop too. SQLiteNode's built-in SQLite keeps every draft, run, and score, and the app updates live while an agent works.

An MCP server + local web app to let Claude Code improve a piece of writing in a loop against natural language criteria evaluated using Jev by TypeSafe AI.

You give it a draft, the context it has to satisfy (a job posting, an essay prompt), and the criteria you care about phrased as statements. The agent rewrites the weakest parts and rankings are shown in the web app. One revision took a resume from 83.2 to 96.6. Each statement gets a 0 to 100 score from Jev, a classifier model that returns scores instead of text and answers in under half a second. Jev is 30x faster and 50x cheaper than Claude Haiku 4.5.

GitHub

Agent Trials

PythonFlask platform on Cloud Run serving every room, the HTTP action API, and the asynchronous benchmark grading.

A platform where AI agents compete over plain HTTP — no runtime to install, just game state in and actions out. Nine rooms spanning games and benchmarks, from Battleship with a weapon economy to code and long-context evals, with Elo matchmaking and browser spectating.

Live

AceAnything

PythonThe FastAPI backend that generates the quizzes and schedules the spaced repetition.

An AI quiz platform I built (FastAPI + Vue + Gemini) that generates adaptive quizzes on any subject with spaced repetition, instant explanations, and text-to-speech.

GitHub Live

Blackjack-On-Chain

SoliditySeven modular contracts hold every rule of the game on-chain.

A decentralized blackjack dApp with seven modular smart contracts running all game logic on-chain, deployed on Arbitrum Sepolia.

GitHub

pi-mcp-server

TypeScriptThe server itself: the MCP tool definitions bridging Claude Code and pi.

An MCP server I wrote to bridge Claude Code and the pi coding agent.

GitHub
More on GitHub →

Books & runs

What I'm reading

  1. Traction — Gabriel Weinberg & Justin Mares (book cover)

    Traction Reading

    Gabriel Weinberg & Justin Mares

    Nineteen channels for getting customers, from the DuckDuckGo founder.

    38%
  2. The Lean Startup — Eric Ries (book cover)

    The Lean Startup Recently finished

    Eric Ries

    Build-measure-learn, validated learning.

    100%
  3. Steve Jobs — Walter Isaacson (book cover)

    Steve Jobs Would recommend

    Walter Isaacson

    The authorized biography.

    100%
  4. Zero to One — Peter Thiel (book cover)

    Zero to One Would recommend

    Peter Thiel

    One of my favorites. Extremely short and sweet. It's a must-read for everyone in the Bay.

    100%
  5. Breakneck — Dan Wang (book cover)

    Breakneck Would recommend

    Dan Wang

    China's industrial policy & manufacturing.

    100%
  6. Nexus — Yuval Noah Harari (book cover)

    Nexus Would recommend

    Yuval Noah Harari

    Information networks, AI, human coordination.

    100%
  7. Business Adventures — John Brooks (book cover)

    Business Adventures

    John Brooks

    Classic Wall Street case studies.

    100%
  8. Blitzscaling — Reid Hoffman & Chris Yeh (book cover)

    Blitzscaling

    Reid Hoffman & Chris Yeh

    Hypergrowth strategy.

    100%
  9. Atomic Habits — James Clear (book cover)

    Atomic Habits Would recommend

    James Clear

    Habit systems.

    100%
  10. The Shock Doctrine — Naomi Klein (book cover)

    The Shock Doctrine Would recommend

    Naomi Klein

    Disaster capitalism.

    100%
  11. Poor Charlie's Almanack — Charlie Munger (book cover)

    Poor Charlie's Almanack

    Charlie Munger

    Mental models.

    100%
  12. The Essays of Warren Buffett — Lawrence A. Cunningham (book cover)

    The Essays of Warren Buffett

    Lawrence A. Cunningham

    Buffett's shareholder letters, by theme.

    100%
  13. Barbarians at the Gate — Bryan Burrough & John Helyar (book cover)

    Barbarians at the Gate

    Bryan Burrough & John Helyar

    The RJR Nabisco LBO.

    60%
  14. A Promised Land — Barack Obama (book cover)

    A Promised Land

    Barack Obama

    Presidential memoir, volume one.

    40%

Runner Frontier — My personal running log

I built Runner Frontier to turn PRs into a continuous frontier. I didn't want my progress to be defined only by arbitrary benchmarks like my one-mile, two-mile, or 5K times. I wanted to measure my improvement over time, in my own way. You can see a summary of my data in the cards below or look at the graph with the raw data here →.

Runner Frontier graph plotting running pace against distance, with monthly run points and the frontier highlighted Runner Frontier preview Pace × distance, run by run Open the live graph
Latest verified snapshot Jan 5 → Oct 2 / 2026
Total distance 266.5miles
Runs logged 121
Time moving 43h 12m

2026 by month

Mileage / runs
Jan17.8 mi10 runs
Feb19.8 mi12 runs
Mar26.0 mi18 runs
Apr32.2 mi15 runs
May27.7 mi11 runs
Jun42.6 mi16 runs
Jul44.5 mi15 runs
Aug31.8 mi13 runs
Sep22.2 mi10 runs
Oct2.0 mi1 run

Personal records

All time
Fastest 1 mi+ 7:02/mi

1.0 mi · Sep 7

Fastest 2 mi+ 8:04/mi

2.0 mi · Oct 2

Fastest 5K+ 9:16/mi

3.3 mi · Apr 15

Fastest 4 mi+ 9:34/mi

4.0 mi · May 2

Longest run 6.1 mi

10:50/mi · Jul 2

First run Jan 5

1.2 mi · 10:04/mi

Most recent Tier 1 2.0 mi

8:04/mi · Oct 2

Latest entries

  1. 2.0 miles
    Time
    16:12
    Pace
    8:04/mi
  2. 2.4 miles
    Time
    19:55
    Pace
    8:29/mi
  3. 3.3 miles
    Time
    31:02
    Pace
    9:31/mi
  4. 2.5 miles
    Time
    22:24
    Pace
    8:56/mi