Graph-grounded answers that cite their sources, gateways that give models tools without handing over keys, voice agents that let you interrupt, and routers that cut the bill. Each one measured, reproducible, and honest when it fails.
Hover or focus a panel to open it. Each links to its own design.
Regulatory filings become a property graph of clauses, amendments and enforcements. Queries traverse the graph and the vectors together, and a reflection loop refuses to answer without source-node citations.
An MCP resource server behind OAuth 2.1 with mandatory PKCE. Tool discovery is filtered by grants, a server-side broker holds every upstream secret, and mutating actions pause for human approval.
A full-duplex WebRTC agent with edge voice-activity detection. The playback buffer clears on interruption, tools run mid-sentence behind a filler, and packet loss is concealed rather than heard.
Golden datasets, Ragas and DeepEval metrics, regression detection against a baseline, and a CI gate that blocks the merge. It replays from recorded cassettes, so CI runs without an API key.
A QLoRA-tuned Llama 3.2 3B served by vLLM, behind a LiteLLM router that escalates low-confidence requests to a frontier model. Most traffic never leaves your own hardware.
Scores are shown against their threshold, and a regression blocks the merge.
34%91%
Vector-only versus graph-grounded accuracy, as claimed in the project brief.
2–4s<500ms
Cascaded HTTP voice bot versus streaming WebRTC turnaround.
A model in production is a hypothesis. I treat it like one: write down what should be true, build the instrument that measures it, record every run so it can be replayed, and let the measurement decide what ships. It's slower for a week and much faster for a year.
An answer without a citation path is a guess. Every claim resolves to a source node in the graph.
graphrag-compliance, retrieval design
The credential broker returns clients, never credentials. No secret ever crosses into a tool argument or a model’s context.
mcp-tool-gateway, architecture
Interruption belongs to the transport, not the model. The playback buffer clears in under 50 milliseconds or the moment is already lost.
voice-agent, latency budget
A score is never a bare coloured cell. “0.71 against a 0.75 bar” is the information; a red square is not.
llm-eval-harness, design system
A step is durable before the next one begins. If the process dies mid-run, the trace still tells the truth up to the last action.
web-agent-navigator, run store
Route on confidence, not hope. The small model answers until it is unsure, and only then does the frontier model get the call.
slm-router, routing policy
Compliance retrieval, agent tooling, voice, evaluation and cost-aware model routing.
Start a conversation