Labs
We build before we advise
Active projects where we test architecture decisions, benchmark real hardware, and publish what we learn. These inform our solutions and our product thinking.
Governed AI Routing Pipeline
How do you route enterprise prompts to the right model without sending PII to the cloud? We built a rules-first router with statistical PII detection and SLM fallback. Three domains (healthcare, finance, telecom), 60 synthetic prompts, full audit trail.
Published findings
This work is part of our exploration while building neurelay.ai, an AI governance platform for enterprise tool access control.
Edge AI on Constrained Hardware
Real-time voice AI running fully offline on commodity hardware. We built the streaming speech-to-text infrastructure missing from the local AI ecosystem — sub-second transcription over WebSocket, with voice activity detection, interruption handling, and multi-turn conversation. No cloud APIs, no GPU, no network required.
Tested on Apple Silicon (M1 Max), Intel x86 (i7 20-core, CPU-only), and NVIDIA Jetson (8GB edge). The pipeline delivers natural conversation with interruption handling across all platforms.
Private Document Intelligence
On-premise document retrieval combining OCR, vector search, and structured retrieval pipelines. Local SLM handles intent detection and classification — no data leaves the perimeter. Testing how far you can push retrieval quality without cloud AI.
Write-up coming soon.
Working on a similar problem?
We build these so we understand the problems deeply. Let's discuss yours.
Start a conversation