Where I've worked

AI Solutions Architect / Senior AI Engineer

Designing and delivering AI systems end to end: business-process analysis, solution architecture, PoC/MVP development, evaluation, infrastructure selection, deployment, and production iteration.

Current focus: AI agents, RAG, long-term memory, speech and meeting intelligence, developer tooling, and LLM evaluation. Products →

whytap · session recall · storm · selected consulting engagements

Led a team of 4 ML engineers and managed a 3-person delivery squad for the Construction AI project. Owned solution architecture, delivery, evaluation methodology, code review, and stakeholder communication across production AI projects.

Selected systems:

• Construction AI — an on-prem document-intelligence platform processing 2,000+ checklist items across 30+ PDFs within a 65 GB VRAM constraint. A separate validation gate achieved ~99.9% accuracy on the high-confidence tier and reduced raw manual workload by ~6×. Case →
• AI Code Reviewer — designed and launched in 3 weeks. During its first month, the system completed 578 reviews, identified 44 confirmed critical defects, and produced 600+ confirmed findings. Case →
• AI Meeting Assistant — transcription, diarization, speaker identification, structured summaries, task export, and Q&A over meeting history.

vLLM · Langfuse · Qwen-VL · Qdrant · E5

Owned pipeline health metrics (ingestion / freshness / context delivery), the evaluation framework, and agentic retrieval tools for an enterprise RAG platform over SharePoint (~12k sources) that unified departmental RAG pipelines. Case →

Azure ML Studio · Azure OpenAI · Azure Search · Ragas

Designed and delivered smart search over product composition with calorie filtering: 76% conversion to product, search time 54 → 35 s. Quantization shrank the model from 30 to 13 GB (generation 1.4 → 0.6 s). 60 RPS on 4× A100.

vLLM · Mistral · quantization · E5

Built the end-to-end agentic support pipeline at 2–6k tickets/day: auto-replies to ~20% of requests (99%+ confidence) cut load by 20%; operator suggestions brought a ticket from 1.5 min down to 37 s. Rolled out to 3 teams. Plus an AI meeting note-taker that files tasks into Jira.

LangChain · vLLM · Qwen · Milvus · faster-whisper