ScholarshipHub

Agent Flow CI

AgentFlow CI is a production-ready, enterprise-grade autonomous software delivery system. It accepts a single plain-English feature request and routes it through five specialized AI agents that communicate, collaborate, loop back, and hand off structured context through Band agent rooms — delivering a fully planned, coded, reviewed, tested, and documented pull request with zero human intervention until the final approval stage. The system addresses a universal bottleneck in software engineering: the time and cognitive load between a feature idea and a PR-ready implementation. AgentFlow CI eliminates that gap by making five AI agents work together the way a real engineering team does — debating, iterating, and handing off work through a shared collaboration layer powered by Band.

DealMaker

DealMaker is an AI-assisted deal desk for B2B sales and commercial approval workflows. It turns raw sales conversations into structured deal records, extracts key details like client name, deal value, scope, decision maker, and contact information, then helps sales teams fill any missing approval-critical fields. From there, DealMaker generates an internal submission for business review, evaluates the proposal against policy, scores risk and compliance, and supports approve or reject decisions with an auditable trail. The project demonstrates how multi-agent AI can move revenue work from conversation to contract-ready approval while keeping humans in control of commercial risk.

Spec6: The enterprise intelligence operator

Spec6 uses Speechmatics in two distinct but connected ways: first as the live speech interface for the product, and second as a transcription engine for “spoken web” intelligence. That split is important, because this is not just “we added voice chat.” In the code, Speechmatics sits both at the front door of the system, where a user talks to Spec6 in real time, and deeper in the intelligence pipeline, where audio and video from the web can be converted into analyzable evidence. The live voice side starts in the frontend voice assistant at [/Users/nadhi/Desktop/win-win/src/frontend/react/chat/voice-assistant.tsx](/Users/nadhi/Desktop/win-win/src/frontend/react/chat/voice-assistant.tsx). The browser captures microphone input with `getUserMedia`, then converts the audio into the format Speechmatics expects: 16kHz mono PCM16. That matters because the browser does not just upload a blob after the user is done. It streams audio continuously in small chunks, which makes the experience feel like an actual assistant instead of a push-to-record form. The comment in that file describes the full loop clearly: mic input becomes PCM16, that stream goes to `/api/voice/transcribe/ws`, partial and final transcripts come back, and the final transcript is fed into the normal Spec6 agent loop. So the voice layer is not a separate AI product. It is an alternate input mode into the same reasoning engine that powers the rest of the app.

Adversarial Infrastructure Analyzer

Every phishing campaign, every business email compromise, every credential-harvesting operation begins the same way: an attacker registers a domain. That moment of registration is the earliest possible signal, and it happens in public. Certificate transparency logs record it. WHOIS feeds publish it. Passive DNS aggregators pick it up within minutes. The entire adversarial infrastructure lifecycle — from registration through weaponization to first victim — plays out across sources that are open, crawlable, and continuous. The core insight is that adversarial infrastructure has a predictable birth cycle — a domain gets registered, gets a TLS cert, gets DNS configured, and only then gets pointed at victims. The window between registration and first use is where you want to catch it, before any SIEM sees a packet. The pipeline works by watching the public signals of that birth cycle in real time.

LongTail Scout

LongTail Scout is a GTM-intelligence agent that surfaces the operators Apollo, ZoomInfo, and Clay can't find: the small, local, niche businesses — roofers, HVAC techs, dental clinics, salons — that aren't on LinkedIn. Those tools are built on the LinkedIn employee graph, so the long tail is invisible to them. LongTail Scout finds these operators through their own websites + Google + news in real time via Bright Data, and proves they're Apollo-blind. Type a niche x city — "roofing contractors in Houston" — and a 3-phase agent runs on a Cloudflare Worker: 1. Discovery — SERP queries via the Bright Data Scraping Browser, rendering Google to bypass bot detection. 2. Enrichment — each homepage and careers page rendered through the Scraping Browser; hiring, size, and recent activity extracted. 3. Synthesis — DeepSeek ranks the operators and writes a sales angle, with every claim citation-linked to the Bright Data fetch that produced it. Each operator's drill-down closes the GTM loop, all Bright-Data-backed: - Apollo-blind verification: a live LinkedIn search confirms the operator has no company page — hard proof Apollo can't see them. - Contact discovery: a real email, phone, and named contact from the contact pages. - Decision-maker finder: the owner/founder plus their LinkedIn profile. - Signal radar: live third-party news (funding, expansion, awards) as buying triggers. - Account brief: a one-click Markdown dossier with every source cited. Niche Recon reverses the funnel: paste what you sell, get the top verticals ranked by "Apollo-thinness" — the share of operators with no own domain Apollo can match — cross-referenced against a private 7M-business demand index. All 12 capabilities are also exposed as MCP tools, so any MCP client (Claude Desktop, Cursor) can drive the whole pipeline.