Engineering Philosophy
"I build AI systems that don't just demo well, but survive production. True engineering is found in the boundaries between the model and the infrastructure—managing asynchronous task queues, ensuring sub-100ms perceived latencies via Server-Sent Events, and rigorously evaluating memory-efficient data pipelines. I believe in local-first privacy, decoupled architectures, and the uncompromising need for robust, reproducible evaluation."
Case Studies
Cortex AI OS
The Flagship (2026)
A 6-service microservices monorepo (Next.js 16, FastAPI, PostgreSQL/pgvector, Redis, Celery, Ollama, Docker) built to solve the privacy compromises of cloud-based LLMs.
Latency Optimization:
Swapped standard polling for Server-Sent Events (SSE) infrastructure, driving real-time LLM response streaming with sub-100ms perceived latency.
Asynchronous Processing:
Delegated heavy document ingestion to Celery workers backed by Redis, ensuring the main FastAPI event loop remains unblocked.
AI Tutor Bot
Stateful RAG (2025)
Engineered multi-turn stateful memory (ConversationBufferWindowMemory, k=5) to maintain deep conversational context without blowing out the token window. Integrated a live DuckDuckGo Search fallback layer to actively ground responses when internal retrieval yields low-confidence answers.
Custom UNIX-Like Shell
Systems Internals (2024)
Developed a shell from scratch using raw C/C++ POSIX calls (`fork`, `execvp`, `waitpid`). Engineered robust command parsing and full support for foreground and background process execution. Proves a rigorous understanding of computer science fundamentals that goes far beyond API integration.
Research Translation
Real-Time Device-Free Human Activity Recognition Using WiFi CSI (AIMLCPS 2026)
The Pipeline:PCA dimensionality reduction (384 down to 64 components, >95% variance), 1D Convolutional WGAN-GP for minority class synthesis, and a hybrid CNN-GRU classifier.
Results: Achieved in-domain Macro-F1 of 77.28% with an end-to-end inference latency of 0.868 milliseconds on standard CPU.
Lessons Learned (The Hard Truth): Testing in a new environment dropped Macro-F1 to 20.99%. This cross-domain failure exposed a structural limitation of amplitude-based CSI, proving that phase-sanitized Doppler extraction is the necessary next step.
Let's Build.
I am actively looking for roles as a Founding AI Engineer or Backend Engineer at early-stage startups and applied AI teams.