Projects

Applied AI systems spanning simulation, retrieval, and product tooling, built with an emphasis on instrumentation, reproducibility, and systems-level iteration.

2022–Present · Ongoing systems exploration

SYS 01IN PROGRESS

Agent Learning Lab

IN PROGRESS

Intent

Make agent learning legible by externalising curriculum and reward design as controllable variables.

System

A scaffolded RL environment where curriculum phases, reward shaping, and observation constraints are explicit knobs rather than hidden training artefacts.

Question

How can curriculum learning and reward design enable a deep RL agent to progressively acquire FPS-like skills?

gdscript · godot 4.5 · python · pytorch · ppo · tcp · ffmpeg · linux

• Godot environment runs as a standalone binary and streams observations, rewards, and done flags over TCP.
• Python PPO loop consumes streamed data and writes structured per-episode logs.
• Replay capture renders frames and encodes MP4s for deterministic playback, directly from within Godot.

Runs

W&B Run →

VERY EARLY · Replay of agent learning to run to a goal

Current: Navigation works with observable learning progress; long-term aim is FPS competence. Blocked on reducing reward-loop wiggle without over-constraining exploration.

SYS 02IN PROGRESS

Litgraph

IN PROGRESS

Intent

Use semantics and knowledge graphs to enrich research and to reveal hidden connections between papers.

System

A research-synthesis prosthetic that provides a graph-augmented retrieval interface over arXiv papers.

Question

Which epistemic relationships are missed by semantic search alone, and how does graph context change retrieval behaviour?

python · fastapi · redis · qdrant · react · vite

• ArXiv discovery feeds a PaperIndex, then embeddings are queued and upserted to Qdrant.
• Vector results are merged with explicit graph context before ranking.
• End-to-end traceability keeps retrieval decisions inspectable and auditable.

Read Writeup →→

pipeline_run · vector_graph_merge · dummy_data

Current: Pipeline works end-to-end; advanced graph reasoning remains planned.

SYS 03IN PROGRESS

Langdeck

IN PROGRESS

Intent

Generate language-learning Anki decks from LLM-generated sentences, with audio and spaced-repetition-friendly formatting.

System

Local-first FastAPI service that uses LLMs to generate Thai Anki decks from prompts with audio and romanisation.

python · fastapi · postgresql · uv · litellm · gcp tts · genanki

• FastAPI service exposes `/generate` and Swagger UI for desktop/mobile access.
• PostgreSQL persistence stores sentences, metadata, and reuse decisions.
• LiteLLM powers Thai sentence generation and duplicate classification.

Coming soon

deck_build · anki_package · audio_cached

Current: Deterministic pipeline works end-to-end locally with Swagger UI; currently focused on Thai.

EXP 01IN PROGRESS

MineGen

IN PROGRESS

Intent

Probe how spatial abstraction emerges in voxel generation and where control breaks down.

System

Early procedural experiment exploring terrain and initial building-generation ideas in voxel space.

Question

What does a hierarchical generator actually learn about spatial structure in voxel environments, and where does controllability break down?

minecraft · python · procedural · deep learning · transformer · convolutional · hierarchical

Coming soon

seed_4512 · terrain_overview

Current: Terrain generation works; building-generation direction remains exploratory and not yet realised.

EXP 02IN PROGRESS

WhereAmAI

IN PROGRESS

Intent

Can we predict where we are from visual cues alone, and can we make those predictions legible?

System

Early-stage interpretability probe aimed at mapping predictions back to concrete visual signals.

Question

Can geolocation be decomposed into identifiable visual cues instead of black-box confidence scores?

python · computer vision · dataset · classification

Coming soon

dataset_build · intake_queue

Current: Conceptual stage; dataset assembly underway, training and interpretability not yet operational. Goal is to train some models and then probe them for which visual cues they rely on for geolocation predictions.

TOOL 01EARLY BUILD

Traypad

EARLY BUILD

Intent

Reduce capture friction as a cognitive offload, without turning into a full productivity suite.

Current: Very early and bare; local capture works, but most features and polish are missing.

kotlin · android · local storage

View Repo →

TOOL 02COMPLETE (HACKATHON PROTOTYPE)

Upkeep

COMPLETE (HACKATHON PROTOTYPE)

Intent

Make local cleanup feel actionable by turning it into a fast, game-like loop of pins, proof, and points.

Current: Hackathon prototype complete: pins, bounty lifecycle, proof uploads, and leaderboard work. AI validation, badges, team challenges, and mobile build were out of scope.

python · fastapi · pydantic · supabase · next.js · tailwind · google maps api · vercel · google cloud functions

View Repo →