.png&w=256&q=75)
🤓 Latest Submissions
PDOS - Private Discovery Oracle System
Organizations frequently suffer from institutional amnesia, where knowledge is trapped in departmental silos and employees are forced to manually hunt for basic information. PDOS (Private Discovery Oracle System) solves this problem by acting as a secure, internal document search assistant tailored for mid-to-large organizations. Powered by a robust Retrieval-Augmented Generation (RAG) pipeline, PDOS allows employees to query internal project documentation and receive immediate answers. What sets PDOS apart is its strict adherence to corporate hierarchy and security: Clearance-Gated Retrieval: All searches are filtered by the user's specific clearance tier (e.g., All Staff, Management, Executive), enforced at the database level using Row Level Security. Human-in-the-Loop Access Requests: If an employee asks a question that requires higher clearance, they are not met with a dead end. Instead, the system generates a live access request that HR or Admin staff can review. Once approved, the employee's answer updates immediately. Comprehensive Admin Controls: Administrators have a dedicated panel to manage employee access, upload and parse new project PDFs directly from the client side, and monitor a complete audit log of all access decisions. Built entirely using native.builder, the platform leverages Supabase (Postgres with pgvector, Edge Functions, Realtime, and Auth) for its secure backend and Fireworks AI (nomic-embed-text-v1.5 and GPT OSS 120B) for its advanced embeddings and chat completions. PDOS ensures that an organization's knowledge is democratized efficiently without ever compromising on data security.
10 Aug 2026
JobShield — AI Job Scam Detection Agent
Job scam losses reported to the FTC jumped from $90M to $501M in recent years, and a third of US job seekers have run into one — yet only 61% feel confident they could actually spot a scam if they saw it. JobShield is an agent that checks a job posting's claims against the real world instead of pattern-matching against old training data. It extracts what a posting claims (company, domain, salary, contact method), then a verification agent decides which tools to call: WHOIS domain age, live search for the company's real domain, a typosquat check, a reviews/complaints search, and a LinkedIn presence check. A deterministic scoring layer combines these hard signals with an LLM pass that writes grounded, evidence-backed explanations — "domain registered 12 days ago," not "this feels off." It works on any pasted text, not just one platform's listings — covering the off-platform pivot to WhatsApp or Telegram that platform-native moderation can't reach. Built on Fireworks AI (a confirmed AMD Instinct GPU partner), Streamlit, and Docker. Benchmarked against a public labeled dataset, and honestly reported even where the result was uncomfortable — a finding that mostly validated why real-time verification matters more than static pattern matching.
13 Jul 2026
