🤓 Latest Submissions
Genie Lobster Bricks
Genie Lobster Bricks is an MIT-licensed deep prompt inspection (DPI) proxy that sits between AI clients and any OpenAI-compatible backend—including Databricks Genie and Databricks model serving. Every user prompt and every model response passes through a programmable firewall: YAML policies enforce intent rules, risk scoring, credential and PII detection, prompt-injection blocks, and data-exfiltration guardrails without changing application code. Inspection uses deterministic regex-based DPI, not an LLM judging an LLM. The project ships adversarial test suites, JSONL audit logs, and optional SIEM export (e.g. Microsoft Sentinel). The demo use case is a fictional Bayou Energy Holdings × Vanguard Well Services Permian completions engagement: contractors use Chatbox pointed at the proxy, not raw Databricks; only three target wells are in scope; reserves, MNPI, and out-of-basin queries are denied with full audit trails. Hosted deployment runs on AWS (API Gateway, ECS Fargate, Secrets Manager) in front of Databricks Genie, with contractor API keys and signed policy bundles. Genie Lobster Bricks is a reusable trust layer for agent guardrails, observability, and compliance-ready AI access control.
19 May 2026
AppBid
AppBid is an AI-native reverse-auction marketplace for auto lending. Instead of sending a full credit application to one lender at a time, I designed it so a dealer submits a PII-free bid request (FICO band, vehicle type, term, amount, state, reserve target). Multiple lender agents then evaluate the same request in parallel and return competitive offers in real time. Each lender is modeled with its own policy profile, and underwriting decisions are generated by a shared Qwen 72B serving stack on an AMD MI300X, with lender-specific behavior tuned through LoRA/policy conditioning. This creates realistic differences in APR, max amount, max LTV, stipulations, and dealer incentive logic by lender. The UI provides a live request/bid stream, lender-level competition, and acceptance flow, plus real-time GPU telemetry and traffic visibility during load simulation. For the hackathon demo, I implemented continuous multi-request generation (“Run demo now / Stop demo”), live bid streaming on the requests screen, and web-console performance observability. The system is designed to show both product value (faster lender competition, better offer discovery) and technical value (single-GPU high-throughput inference, structured decisioning, and deployable open-stack AI infrastructure).
10 May 2026
