Projects
Production-grade AI systems, multi-agent pipelines, RAG architectures, and ML projects.
Agentic AI system that autonomously parses complex simulation logs, identifies anomalies, and summarizes key performance metrics using CrewAI and LangChain.
Reduced manual review time by 75%
Multi-agent platform where autonomous agents collaborate to perform real-time evaluation of business and technology proposals — Flask backend, React frontend.
Full-stack AI-powered decision support
Agentic pipeline that ingests city email notifications, scrapes city websites, extracts meeting details, and generates structured calendars, agendas, and summaries.
1M+ records/week · ~30% accuracy improvement
Hybrid retrieval system combining Neo4j knowledge graphs with vector databases (Weaviate, FAISS) for multi-hop reasoning and context-aware search.
Improved retrieval relevance by ~30%
ML pipeline for an AI-driven hardware pet health product, integrating on-device firmware with cloud backend and processing high-frequency IoT telemetry.
Real-time pet health anomaly detection
Web app to automatically scrape and process transcripts from city meeting videos, enabling downstream NLP and summarization pipelines.
Feeds 1M+ records/week pipeline
Internal Python + Shell data pipelines to automate inventory tracking, sales reporting, and analytics dashboards — driving 30% YoY product growth.
30% YoY growth · 60% less manual reporting
A systems-focused project exploring how local LLMs behave under different memory, inference, and fine-tuning conditions — GPU memory allocation, quantization, context length, and benchmarking across models.
VRAM, tokens/sec & load-time benchmarks across models
Low-level experiments on GPU memory behavior, CPU–GPU transfer costs, and runtime performance — allocated vs reserved VRAM, pinned memory, async copies, streams, and CUDA event timing.
A practical view of what happens below the model API
A benchmarking harness comparing local LLM inference across model sizes, quantization modes, and context lengths — tracking tokens/sec, peak VRAM, model load time, and output quality.
Disciplined benchmarking over subjective impressions
A fine-tuning project adapting a local model with QLoRA while tracking VRAM, training loss, and before/after output quality on a CUDA & quantization instruction dataset.
Lower loss ≠ better answers — measured behavior change
A systems simulation of how multiple workloads share a GPU — comparing FIFO, priority, round-robin, and deadline-aware scheduling against p50/p95/p99 latency, throughput, and deadline misses.
Predictability & tradeoffs in shared compute
A computer-vision pipeline measuring latency from capture to GPU processing — frame capture, CPU preprocessing, CPU→GPU transfer, inference, and end-to-end frame latency / FPS.
Stage-by-stage latency in real-time perception
A production website designed and developed end-to-end for a local automotive business — services, trust signals, and contact, from requirements through deployment and maintenance.
Live customer-facing site
A full-stack communication platform (FastAPI, Next.js, MongoDB) for scalable outreach — contact management, API integrations, and reliable message-delivery workflows.
Scalable outreach & message delivery
A document search and retrieval system for natural-language questions over large collections — combining vector search, metadata filtering, and LLM summarization behind backend APIs.
Natural-language search over large document sets
More experiments, research notebooks, and open-source work on GitHub.
View GitHub Profile