PROJECTS

↳ How it fitsOne semantic layer defines every metric exactly once for both engines. The deterministic parser and the LLM planner emit the same plan schema, checked one key at a time before the guarded compiler writes any SQL. Your key stays in your tab. Calls go straight from the browser to the provider.

↳ How it fitsEach agent lives in its own LangGraph node and talks to Claude, GPT, Gemini, or Grok. They really do disagree. A moderator weighs the arguments and turns the spread of opinion into the confidence score. The disagreement is the metric.

↳ How it fitsThe less glamorous platform work keeps the fleet alive. The scheduler handles cron, jitter, and ramp-up. Warmed persistent identities come from a pool, health-scored proxies disable themselves, failures are classified with alerting, and costs are metered against a hard ceiling that stops a runaway run. Every action must match a strict JSON schema, so the model can only click, fill, scroll, or stop. I can VNC into any live worker and watch it think. Postgres and Redis hold it together, unattended, on a single Mac Mini in the corner.

↳ How it fitsTwilio streams in the call audio and Whisper turns it into text. An agent works out what to say, checks the calendar for an open slot, speaks the answer, and writes the booking. It all moves fast enough to feel like a normal conversation.

↳ How it fitsI cleaned the data in pandas, built an explicit one-hot feature matrix, and tested logistic regression against gradient boosting. The readable model wins, with an AUC of 0.845 versus 0.830. Each prediction is the sum of named coefficients, so the live demo shows exactly why a customer got that score. It is the model's own arithmetic, not a post-hoc explainer.

↳ How it fitsIt calls kubectl for pods, logs, and events, matches the evidence to vetted failure signatures, then drafts a fix from templates with a risk grade and staged rollback. The approval gate lives in code, not convention. The subprocess layer refuses mutating commands until a human hits approve.

↳ How it fitsIt captures the call audio and the rep's screen, then transcribes locally with almost no lag. That context streams to a model, which drafts the next answer or step for a glanceable side panel. New team members get live guidance without breaking the flow of the call.

↳ How it fitsA Go orchestrator schedules one Docker sandbox per agent task on Kubernetes and streams stdout and results back over NATS. Redis tracks state, so nothing leaks from one run to the next.

↳ How it fitsLangChain splitters chunk the runbooks, incidents, PDFs, CSVs, and diagram notes, then they are embedded into Postgres with pgvector. A hand-rolled hybrid ranker combines cosine similarity and keywords behind a relevance floor. Groq drafts the cited answer quickly, and Redis caches hot queries for instant repeats. The live demo runs the same pipeline entirely in your browser with real embeddings and no server.

↳ How it fitsThe staging-and-facts pipeline really runs, with SQL and data-quality checks after every task. Diagnoses use deterministic signature matches over probed evidence, and fixes come from vetted templates. Destructive repairs snapshot the table first. The demo replays a genuine end-to-end run: the data actually broke, and it actually got repaired.

↳ How it fitsThe whole brief runs client-side. The browser builds the dollar scoreboard and calculates the debt's pace in dollars per second, clearly labeled as an extrapolation between daily Treasury postings. It also handles the math for each resident. Every claim carries its source and fetch time. If a feed is down, Pulse says so. The same engine ships as a Python CLI with offline fixture tests.