dbinfer

LLM Inference Engine. C++23, Metal, NEON

github.com/druhinb/dbinfer

• Built an Apple Silicon LLM inference engine from scratch in C++23, writing bespoke CPU (NEON) and GPU (Metal) kernels.
• Achieved 147 tok/s on M3 Max (Qwen2.5-0.5B Q8_0), 60% of llama.cpp's Metal backend and up from an unoptimized 26 tok/s.
• Implemented speculative decoding with verification on GPU, achieving a 1.94x speedup and identical output to the target model.

drubuntu

ARMv8 Multikernel OS. C, ARMv8

github.com/druhinb/drubuntu

• Built the memory system from bare metal: a buddy allocator managing RAM in 4KB pages and a four-level page-table walker.
• Implemented same-core and cross-core message passing between processes, including UMP, a lock-free shared-memory channel.
• Wrote the ELF program loader, per-process address spaces, and multicore bring-up to run user processes across all 4 cores.

Vulnerability Analyst

NVD Chatbot. Python, LangChain, llama.cpp

github.com/druhinb/Vulnerability-Analyst

• A ReAct agent in the terminal that answers questions from the National Vulnerability Database.
• Three tools: look up a CVE with its CVSS scores, search vulnerabilities by keyword, and read the advisories a CVE links to.
• Runs on OpenAI or on a local GGUF checkpoint through llama.cpp, picked with one flag.

snowfight

4X Strategy Game. C#, SDL2

github.com/druhinb/snowfight

• A local multiplayer 4X in the spirit of Civilization I, on a small SDL2 engine, with snowball fights for combat.
• Maps come from Perlin noise with biomes, features, and productivity per tile, tuned from a menu before the game starts.
• Factions found cities and towns, climb a tech tree, and field kids, teens, and adults; games save to disk between sessions.

Shape of Thought

Probing Self-Supervised Vision Models. Python, PyTorch, Hydra

github.com/druhinb/LatentInvestigation

• Asked how much 3D pose DINOv2, I-JEPA, MoCo v3, and a supervised ViT keep in their features, using ShapeNet renders from 24 viewpoints.
• Trained linear and MLP probes to predict azimuth and elevation from six layers of each model, then compared the layer curves.
• Added a voxel reconstruction head on frozen features to test how much shape survives, with every run configured through Hydra.
• Found viewpoint and shape live in the middle layers of the self-supervised models, where the supervised ViT falls behind.

Roblox

Software Engineer Intern. June 2026 - Sept 2026. San Mateo, CA

• Shipped server logs and post-mortem server history to the creator dashboard; 60% of top 500 creators adopted them within a month.
• Built Flink pipelines ingesting combined 2M events/sec of client and server logs from Kafka into sharded ClickHouse.
• Owned the read path, designed Elasticsearch cluster to serve server details in <200ms; ClickHouse logs in <500ms.
• Designed container connectivity, compute scheduling, and queue trigger design for Roblox's from-scratch serverless compute platform.
• Built serverless worker in Rust & C; gVisor isolation, custom Luau runtime, SPIFFE mTLS, sub-100ms in-game invocations.
• Won signoffs and SVP approval for serverless designs and capacity estimates across 7 partner teams, from engine security to infra.
• Halved payload sizes via Brotli compression and MessagePack in Cross-Server Messaging, Roblox's C++ inter-server bus.

DropzoneAI

Research & Development Intern. Sept 2025 - Dec 2025. Seattle, WA

• Designed long-term agent memory on pgvector that carries context from past investigations to cut median investigation time 35%.
• Built an eval from real customer incidents and CVEs, supporting dynamic model routing and benchmarking for new features.
• Raised percentage of alerts resolved autonomously from 70% to 90% by developing a judge harness to screen sensitive actions.
• Architected a distillation pipeline to compress info from EDR telemetry, and alerts by 80% with no loss in investigation efficacy.

Amazon Web Services

Software Development Engineer Intern. June 2025 - Sept 2025. Seattle, WA

• Built new shard-failover pipeline for Kinesis Data Streams (Tier-0, 99.99% SLA), retiring a recurring class of long-tail outages.
• Modeled hardware telemetry in a Java/Rust service to proactively migrate at-risk shards, cutting heat-related disruptions 24%.
• Added real-time host-failure detection over fleet telemetry, triggering shard failover before customer latency degraded.
• Drove the rollout from canaried telemetry streams to every us-west-2 host across heat, storage, and control-plane owners.

DevMatch

Software Engineer Intern. Oct 2024 - June 2025. Seattle, WA

• Built a browser IDE on AKS to sandbox 20,000+ untrusted code runs; per-tenant isolation, warm-pooling, and DinD sidecars.
• Instrumented the platform with OpenTelemetry, as well as standing up Prometheus alerting and Grafana dashboards on top.