Second year computing science at the University of Alberta. I write low level systems
from scratch to find out what is underneath the abstraction: an allocator instead of
malloc, Raft instead of a database that already handles consensus, a rasterizer
instead of a graphics API.
None of it is better than the real thing. That was never the point. The point is that afterwards the real thing stops being magic, so when it behaves strangely you have somewhere to start.
A fully local autonomous voice assistant. Sub-600ms bidirectional speech with barge-in, OS level machine control, dual layer persistent memory over a knowledge graph, and a WebGL holographic HUD. It boots offline with no GPU, no model download and no credentials, then swaps in real backends behind traits.
oracle-audio (C++/RT) ──shm+socket──▶ oracle-core (Rust/Tokio) ──WS──▶ oracle-hud
capture · VAD · agent loop · memory ·
barge-in · TTS connectors · gateway
│
authed UDS (SO_PEERCRED)
▼
oracle-actd (Rust, privileged)
policy · input · shell · audit
The design premise is that the model is an untrusted planner. Actuation lives in a separate privileged daemon that recomputes the required capability from the operation itself, gates irreversible actions behind confirmation, and audits everything.
github.com/apollo-2006/oracle-of-delphi
Open source work on ggml-org, the C/C++ inference stack behind whisper.cpp and llama.cpp.
whisper.cpp #4031  |
the tests aborted on any GGML_BACKEND_DL=ON build: no backend registers until something calls ggml_backend_load_all(), so init ran with zero devices and tripped an assert. every example already made that call, the tests were the only callers that did not. found it, fixed it in four call sites |
| whisper.cpp #4029 | prebuilt macOS CLI binaries: arm64, x64, and a lipo fused universal archive that picks its CPU backend at load time, so one download runs on an M1 and an M4 without either giving up its instructions |
| whisper.cpp #4019 | documented the stream example's two output formats, and the CWD relative model path that surfaces as a context init failure |
| llama.cpp #28261 | documented streaming tool call deltas and the Jinja requirement, every claim cited to file and line |
| nexus_db | embedded LSM key-value store: skip list memtable, write ahead log, immutable SSTables, tombstone compaction |
| nexus_cluster | Raft from scratch: leader election, log replication, term management, custom TCP transport |
| nexus_editor | real time collaborative editor, fractional indexing CRDT, goroutine per client |
| nano_match | limit order book with price time priority and no heap allocation on the matching path |
| custom_mem_alloc | thread safe allocator over one mmap region, intrusive free list, block recycling |
| neon_vm | stack based virtual machine, bytecode chunking, its own dispatch loop |
| cpu_rasterizer | full 3D pipeline in plain JavaScript, no graphics API. barycentric fill, 1/w z-buffer. runs in your browser |
| photon_tracer | raytracer with no libraries: vector math, ray generation, camera, PPM output |
| rasterizer_engine | software rasterizer, perspective matrices and triangle fill written out longhand |
| thermal_monitor | daemon tracking CPU load, thermal curves and VRAM clocks into a local database |
| terminal_dashboard | system monitor with per core CPU, memory, network and GPU telemetry |
| cal-cli | macro tracking from the command line, because logging a meal should be one command |
| points-sys | two player points economy, one HTML file, no build step, live synced |
| valo_scout | valorant stat tracker |
| radiant_slice | barebones valorant style fps shooter |
| personal_portfolio | abirdeol.tech, built from scratch |
Four of the projects above had a bug that cost me real hours. I wrote each one up afterwards: the symptom, how I chased it, what was actually wrong, and what I changed about how I write that kind of code.
- The block that was too small to free an intrusive free list corrupting the block next door
- The delete that did not delete a flush optimization that resurrected deleted keys
- The order that was in two places a use after free that never crashed
- A cluster that could not keep a leader when the runtime pauses longer than your failure detector waits
Working through machine learning from classical computer vision upward, and writing down what I learn rather than collecting tutorials. Two time national MMA champion, 2020 to 2024. Looking for software engineering internships.