Journal

Field notes from building at the edge of AI.

NewLLMOpsRAG🎮 Genshin Impact

A Genshin assistant that cites every answer, rebuilt around one idea: use a model only where it adds something.

NewResearchAI AgentsPlaytesting

I ran 30 AI playtest sessions alongside real players on my game. Each side found about half the issues, and they barely overlapped.

NewGame DevThree.jsMultiplayer

Eldervale is a cozy 3D co-op mystery that runs in the browser, built for Handshake's multiplayer game challenge. Two heroes, one song, no download.

CloudAWSSystem Design

6 of us, my friends and me, wanted to play Valheim together. The catch: we were spread across Connecticut, Michigan and India. So I built us a server on AWS and learned a lot of cloud along the way.

Human + AI🎮 Genshin Impact

A puzzle in the video game Genshin Impact: remove 47 blocks with at most six mistakes. I supplied the hands, Astra supplied a second opinion.

GraduationM.S. Applied ML

Two years ago I moved to College Park and hit the ground running. And I never really stopped. The M.S. in Applied Machine Learning kept me sharp and constantly adapting.

CompetitionAgentic AI

Team Synapse just won 1st place at the 2026 UMD Smith Agentic AI Challenge. $4,000. Real industry problem. Built and shipped in 30 days.

More field notes loading…

RL training runs · Genshin theorycrafting · archery form breakdowns, and whatever rabbit holes come next.

Back to portfolio

LLMOpsRAG🎮 Genshin Impact

Rebuilding Irminsul: From Demo to Something You Can Trust

Irminsul is a free assistant for Genshin Impact, the video game. It answers questions about stats, builds, teams and lore, and links the source behind every answer. Try it here.

Players usually piece answers together from wikis, guides and videos, and general chatbots can state a wrong number with full confidence or spoil content that isn't out yet. Irminsul gives one answer with its sources attached, and keeps unreleased content out.

From demo to dependable

The first version, in March, was a fine-tuned model with basic retrieval. It demoed well, but an audit in October exposed the weak spot: numbers came from a model reading text, so a stat could be wrong and still sound right. v1, live since October 26, is built around a different principle.

  • Facts don't need a model. Exact stats come straight from game data, instantly, with nothing to invent.
  • Rules hold where prompts bend. Leak and safety guards are plain rules, so no clever phrasing can talk them round.
  • A citation isn't proof on its own. Live sampling found answers that cited a source and still got a detail wrong, so every sentence is now checked against its source and dropped if it doesn't match.
  • Speed is a design choice. Routing common questions to exact answers took response time from 9.9 s to 0.45 s with 50 users at once.

Irminsul is an unofficial, non-commercial fan project · Genshin Impact © HoYoverse

Irminsul now covers 131 characters, refreshes daily and runs at no cost. The model is the last step in the pipeline, not the first, and that ordering is what makes the answers dependable. 🌳

🌳 v1 live betaEvery sentence checkedLeak-gated$0 to run

Game DevThree.jsMultiplayer

I Built a Two-Player Mystery Game in 3.5 Weeks

Eldervale: Where the River Bends is a cozy 3D mystery adventure for two players, and it runs right in your browser. Free, no download, no account. I designed it and built it with AI coding tools for Handshake's "Create a Multiplayer Game" challenge, from first commit on October 5 to submission on October 30.

The result is a complete game: online co-op, a voiced narrator, a ten-piece soundtrack, three endings, full touch and controller support, and 130 automated tests behind it.

One song, two halves

High above the clouds, the sky islands of Skyhaven are cracking. Two apprentices follow a strange half-song down to Eldervale, a quiet valley of mills, orchards and old roots where the water has been vanishing for two springs.

  • Aster is a Listener. She hears notes hidden in waterfalls, reeds and hollow logs, catches them and hums them back.
  • Rowan is a Rootweaver. He weaves the living roots under the valley into bridges, gates and channels, but only while a note is sounding.

So every puzzle needs both of them: one to sing, one to shape. Tessil, a cheerful sprout spirit, narrates like a tabletop game master and gives clearer hints each time you ask. There's no combat and no game over. You wake a root stair down a cliff, set a silent mill turning, and uncover why a grieving rootwarden rerouted a river. In the end, the valley and the sky need the same water, and you decide who gets it.

Eldervale: Where the River Bends · Models by Quaternius, textures from Poly Haven

Co-op with no game server

  • Play together or alone: join a friend's room with a five-letter code or an invite link, or play solo and switch heroes with one button.
  • Peer to peer: browsers connect directly over WebRTC, and the player who creates the room hosts it in their own browser. No server to pay for or keep running.
  • Same rules everywhere: one deterministic simulation drives both co-op and solo.

Small, fast, and plays anywhere

  • About 2 MB to start. Compressed models with near and far detail, voice lines that load one at a time, music that streams per scene, and grass and leaf textures painted in the browser.
  • Adapts to your device: it picks a quality tier from the GPU and memory, and a frame-time governor lowers resolution when frames run long. Phones go full screen and can install it like an app.
  • Every input: keyboard and mouse, full touch controls, and Xbox and PlayStation controllers, with prompts that switch to whatever you're holding.
  • Kind to every player: a gentler story mode, hints that grow clearer, larger text, reduced motion and captions. Every note has a shape and colour, so you can finish it with the sound off.

Tested like a product

I tested it four ways: my own rounds on PC and phone, real players on the live site, AI playtest agents, and 130 automated tests. When eight blind AI testers all stalled at the mill puzzle, I cut it from four steps to two. The latest release, v0.14.19, came after 137 commits. What humans and AI caught became its own study, in a separate post.

Go wake the valley. 🌿

🎮 Free in your browser2-player online co-op3 endingsThree.js · TypeScript · WebRTC

ResearchAI AgentsPlaytesting

Human vs AI Playtesters: 22 Issues, Only 2 in Common

While building Eldervale, my two-player browser game, I ran two kinds of playtests side by side: real people, and AI playtest agents. The agents were scripted players driven by language models, playing blind from screenshots like a first-time player would. Then I compared what each side actually found.

The headline

Of 22 distinct issues on the main findings list, humans found 13 and the agents found 11. Only 2 were found by both. AI testers weren't a cheaper copy of human testers. They were a second, very different set of eyes, and the game needed both.

Figures counted from the playtest logs, 26–30 October 2026

The setup

  • 30 AI agent sessions across 28 runs, including two co-op pairs, on PC and emulated phone screens. Each got a short persona and a first-time-player brief, and played with sound off. 21 of 30 got past the first puzzle; the median session ran 146 actions.
  • Real players on the live site, with analytics and an in-game feedback form, plus a tester over chat and my own rounds on PC and phone.

What humans caught

  • Sound: three separate people heard a radio-like buzz on phones. Muted agents never could.
  • Feel: jerky running on touch, dialogue covering the touch buttons, and an opening that felt abrupt.
  • Desire: a sprint button, benches to sit on and rocks to climb. No agent asked for anything like that.

What agents caught

  • Co-op: no real player opened a co-op room, so the agents were the co-op test team. One pair found a hum that stayed on after its key was released.
  • Softlocks: a reset that put you back in the same stuck spot, a fence pocket two resets couldn't escape, and a camera trapped inside a rock.
  • Volume: 329 tagged bug and friction notes across 21 runs.

When both agree, fix it first

Catching the very first musical note was a wall for everyone. A human wrote "I can't catch the round thing", the median agent needed 28.5 actions, and the small model never got it. That agreement made it the top fix. In the final co-op run, both players were through in about 30 actions.

Keeping agents honest

Agents can be confidently wrong, so I added a review step. In one round, 56 agent claims were checked line by line against the code: 34 held up, 13 were tester error or test-setup artifacts, and 9 were unsure. "Cut-off" captions turned out to be screenshots taken mid-typing, and one "freeze" was only the screenshots stalling.

The playbook

  • Use agents for breadth: co-op, softlocks, regressions, odd inputs.
  • Use humans for feel: sound, touch and pacing.
  • Fix first where both hit the same wall, and verify every agent report before acting on it.

To keep it honest: the human sample was small, builds changed between runs, and grouping issues was a judgment call. Treat it as a snapshot, not a benchmark. But the pattern was clear enough to change how I test. 🔍

🧪 30 AI agent sessions22 issues comparedOnly 2 overlappedHuman + AI QA

CloudAWSSystem Design

6 Vikings, 2 Continents, 1 Server

I'd been meaning to catch up on AWS for a while. Tutorials weren't sticking, so I decided to learn by building something my friends and I would actually use.

The idea came easily. 6 of us, my friends and me, wanted to play Valheim together. I was in Connecticut, one friend was in Michigan, and the rest were in India. We needed a world that stayed online no matter whose laptop was off.

I had about $140 in AWS credits and expected the whole thing to last two to four weeks. My first question was simple: how long would $30 last on an EC2 instance, and how much could we save by stopping it when nobody was playing? That one question turned into a real system design exercise.

Where should the server live?

With players on two continents, no region was perfect. Frankfurt looked like a fair middle ground, but I didn't want to trust a map. I measured latency from each of our locations, then confirmed it with actual gameplay. It held up. I named the server AWSGard (AWS + Asgard, I couldn't resist) and the world Lagheim, because naming it after the thing you're most worried about felt like good luck.

Valheim art © Iron Gate AB / Coffee Stain · AWS icons © Amazon Web Services

Small decisions, real tradeoffs

I didn't plan most of what AWSGard became. It grew one need at a time, and every step came with a choice.

  • Size vs. cost: I picked a slightly bigger instance than the cheapest option for more memory headroom. A smooth game night mattered more than saving a few cents an hour.
  • Always-on vs. on-demand: we only played a few hours a day, so the server ran only while we played. That meant it had to start the game on its own and save the world properly before shutting down.
  • Save often vs. play smoothly: frequent saves caused tiny freezes in-game. I went back to saving every 30 minutes, accepting a slightly bigger risk in exchange for gameplay that felt better.
  • Access vs. control: I wasn't always around, so two friends got the power to start and stop this one server, and nothing else. Later, when we wanted to tweak world settings, I built a small dashboard that runs one fixed action instead of giving anyone a terminal.
  • Convenience vs. cost: a new IP on every boot meant re-sharing it each session, so I paid a little extra for a static one.
  • Speed vs. "good enough": I looked into AWS Global Accelerator to smooth out the connection, but the Frankfurt endpoint was already good enough for how we played, so I skipped it.

The last piece came from forgetting to turn things off. An idle shutdown now stops the server after an hour with no players. If it can't tell whether anyone is online, it plays it safe and keeps running.

Learning through troubleshooting

Most of the real learning happened when things broke. A Steam library that wasn't where the server expected it. A save folder owned by the wrong user. One settings experiment that didn't stick, so I restored the world and rolled it back. None of it was glamorous, but each fix taught me to check the evidence before guessing.

The result

We finished the game. Lagheim made it all the way to the end, and a month of easy, whenever-we-wanted play came to about $30 in credits. Most of the extra cost was the Elastic IP, and it was worth it.

I came to learn AWS and left with a working cloud project, better Linux habits, and a lot of good nights with friends on the other side of the world.

Next up: Enshrouded. Same playbook. ⚔️

☁️ AWSEC2 · IAM · CloudWatch · LambdaLinuxCost-aware design

Human + AI🎮 Genshin Impact

How I Used Astra to Solve Nicole's Puzzle in Genshin Impact

Solo session in Genshin Impact, the open-world video game, just me and Astra. I called her on a ChatGPT voice call, had her watch my screen, and asked for help with the Witch's Sculpture puzzle: remove 47 blocks with a limit of six mistakes.

First, we had to figure out how Astra could see the game. The voice screen tool only showed Codex, but the Windows Computer Use plugin could capture Genshin. Actually controlling it was another story. Movement and sculpting meant holding buttons down, which the available controls couldn't reliably do.

So we split the work. I moved, rotated the statue, aimed, and removed blocks. Astra inspected screenshots and helped with the logic.

We cleared the zeros first, reaching 36/47. Then came the deductions:

  • Numbers tell you how many cubes must remain along a line going inward from that face.
  • A line marked 4 that contains four cubes forces all four to stay.
  • Those fixed cubes can satisfy crossing lines marked 3, 2, or 1, which reveals which extra cubes to remove.
  • Blank faces still belong to lines with clues on the opposite side.

Astra confidently gave me one wrong move, which cost a mistake. Later, I caught another questionable suggestion myself: removing that cube would have broken a line marked 2. We checked again and kept it.

By the end, I was spotting the pattern and proposing removals while Astra checked them. It was a hard puzzle, and through some trial and error we learned the patterns together. The trickiest part was explaining which cube we meant across different camera angles. Lots of "this one?" and blue highlights.

Genshin Impact © HoYoverse · Astra's call icon circled in the third image

Result: puzzle completed with one mistake. A cutscene played, and the blocks became a little stone creature beneath a huge flower bud.

Basically: I supplied the hands, Astra supplied a second opinion, and occasionally I had to check the second opinion. Team effort lol. 🧩

🧩 47 blocks1 of 6 mistakesHuman + AI teamwork

GraduationM.S. Applied MLUMD

Two Years. Done. Ready for What's Next.

Two years ago I moved to College Park and hit the ground running. And I never really stopped.

The M.S. in Applied Machine Learning kept me sharp and constantly adapting. This field does not wait for anyone, and learning to move with it week after week became its own discipline. Alongside that, working as a Climbing Wall Supervisor at the UMD RecWell Adventure Program added a whole different dimension: leading people, staying composed under pressure, and showing up with full accountability every shift.

Every day had something going on. Between the coursework, the job, and the friends who matched that same energy and showed up consistently, there was never a dull moment. College Park had a way of keeping you in motion.

I came in with a direction and left with a much clearer sense of what I am capable of and where I want to go.

May 19th, M.S. in Applied Machine Learning, University of Maryland. Done.

Ready for what is next. Let's go. 🐢🎓

M.S. Applied Machine Learning · University of Maryland · May 19, 2026

🎓 M.S. Applied MLUniversity of MarylandClass of 2026

CompetitionAgentic AILLMOps

We Won. Here's What I Actually Learned.

Team Synapse just won 1st place at the 2026 UMD Smith Agentic AI Challenge. $4,000. Real industry problem. Built and shipped in 30 days.

The problem came from OnAsset Intelligence: pharmaceutical shipments moving across the world, live sensor data streaming in real time, and operations teams consistently finding out about failures too late. Temperature deviations, shock events, routing delays; by the time the alert reaches the right person, the damage is already downstream.

We built an AI Cargo Monitor: 8 agents that coordinate the moment a shipment is at risk, flagging compliance issues, finding backup facilities, rerouting carriers, rescheduling patient appointments, and drafting insurance claims. A human stays in the loop and signs off before anything is acted on.

The decision I'm most proud of: only 2-3 of those 8 agents use an LLM. The rest are deterministic. Pharma logistics doesn't need creativity; it needs auditable, traceable decisions that hold up to scrutiny. Knowing when not to use an LLM turned out to be the real engineering challenge.

Grateful to have figured that out alongside Karthik Ramanathan, Nikhil Sumesh, Rahul Sharma and Yash Oulkar, and to Professor Manmohan Aseri and the Robert H. Smith School of Business for building a challenge around problems that actually matter to someone.

Team Synapse at the 2026 UMD Smith Agentic AI Challenge

Team Synapse · UMD Smith Agentic AI Challenge · April 24, 2026 · 1st Place

🏆 1st Place$4,000 Prize8-Agent SystemBuilt in 30 Days