Nexum

Landing Page - Hero Section
client dashboard
client workspace
client payment section
Bryntum Gantt Chart
Elastic Search
Channel 3
freelancer dashboard
freelancer workspace and proof of execution checks
freelancer proposal section
audit logs with ag grid table

Inspiration

Anyone who has worked as a freelancer or hired one knows the quiet dread that creeps into a project mid-way through. It usually starts innocently: a quick message in Slack saying, "Can we also support multi-currency checkouts?" or "Just a tiny tweak to the database schema."

Before long, a two-week sprint turns into a two-month slog. Freelancers end up doing unpaid work out of fear of getting ghosted or receiving a 1-star review. Clients worry they’re being billed for incomplete features or code that breaks in production. Traditional platforms don't solve this-they just hold funds in a black box and hand down arbitrary dispute verdicts when things go sideways.

We wanted to fix this fundamental breakdown of trust. We realized that AI shouldn't replace human agreements or make unilateral decisions. Instead, AI should act as an impartial mediator: drafting specifications, catching verbal scope changes before they snowball, and objectively verifying technical work-while humans retain full signing authority and PayPal guarantees financial liquidity.

What it does

Nexum is an AI-mediated freelance workspace and autonomous escrow settlement platform. It sits between client and freelancer across the entire project lifecycle:

Dual-Boundary Private AI Counsel: Both the client and freelancer have their own dedicated AI assistant. The client’s AI turns rough ideas into detailed, milestone-based PRDs. The freelancer's AI helps structure estimates and technical deliverables. Crucially, private negotiation ceilings and personal notes are isolated at the database query layer-neither party's AI can see the other's private context.

Mediated Shared Workspace & ADRs: In the shared project space, a Tier-1 AI mediator monitors communication (including sync calls via our live AI Notetaker). When scope expansions are discussed, the system automatically detects the delta, calculates the impact on timeline and budget, and generates a formal Architectural Decision Record (ADR). Nothing changes until both parties explicitly ratify it with a dual cryptographic signature.

Autonomous Deliverable Verifier: When a freelancer submits code, the client doesn't have to guess if it works. Nexum runs automated sandbox inspections: AST semantic analysis, mutation testing, dependency supply-chain auditing, adversarial security scans, and Playwright visual assertions, sealed with SLSA v1.0 provenance attestation.

PayPal Milestone Escrow Settlement: Escrow funds are locked via PayPal Orders v2 before sprint authorization. Once deliverables pass verification and both parties approve, funds are released. A 72-hour inspection window ensures clients have time to review, while protecting freelancers from being ghosted.

Mobile Workspace via Telegram: Users can track milestone statuses, digitally sign scope amendments, and approve escrow releases directly on the go through our paired Telegram bot.

How we built it

We built Nexum as a full-stack, AI-mediated freelance platform designed to eliminate freelance disputes through bilateral consensus and human-in-the-loop governance.

Dual AI Mediation & Speech Architecture Google Gemini (AI Copilot & Evaluator): Powers private AI mediators for both client and freelancer. When scope changes or add-ons are requested, Gemini dynamically analyzes the requirements, estimates timeline adjustments, computes fair price deltas, and drafts formal amendment cards. Groq Whisper: Enables zero-friction voice messaging via mobile, instantly transcribing spoken audio into structured workspace updates. Human-in-the-Loop Safeguards: AI outputs remain non-binding drafts until explicitly approved by human counterparties. Nexum uses a hybrid AI architecture combining Google Gemini (Multiple models for fallback system) and DeepSeek models (properly distributed according to complexity of work )
Frontend & Real-Time Collaboration Next.js & React (App Router) with Tailwind CSS: Responsive, high-performance UI tailored with role-specific views (Client Commander vs. Freelancer Specialist). WebRTC Peer-to-Peer: Custom Node.js signaling server supporting low-latency video and audio meetings directly inside the workspace without external meeting links.
Backend, Security & Integrity Node.js, Express & TypeScript: Strict type-safety, endpoint validation using Zod, role-based access control (RBAC), and strict resource ownership verification. Cryptographic Protocol Timeline: Every contract modification, scope addition, deliverable submission, and status transition is sealed with SHA-256 hashes in an append-only audit trail. Escrow & Ledger System: Financial transactions and balances are tracked in integer minor units (zero floating-point math) integrated with payment gateway APIs.
Omnichannel Mobile Telemetry (Telegram Bot API) Built a full-featured Telegram Bot Command Center that acts as a mobile remote for active contracts. Features native file delivery, inline portfolio switchers, real-time Kanban task management, and an intelligent relay dispatcher supporting @mentions and Swipe-to-Reply to route messages directly between Telegram and the web workspace.
Database & Cloud Storage PostgreSQL (Neon Serverless): Relational state management for users, multi-tier contracts, milestones, and audit trails. S3-Compatible Object Storage: Secure storage for client deliverables, attachments, and signed ADR files.

Challenges we ran into

Financial Integrity & Idempotency: Integrating PayPal Orders v2 in an escrow model where funds need to be authorized, captured, and programmatically tied to granular milestones was tricky. We had to ensure race conditions couldn't double-capture or prematurely release funds if someone refreshed or if a network blip occurred. We solved this with rigorous database-level locking, explicit state machines (NOT_STARTED → CAPTURED → RELEASED), and canonical SHA-256 audit hashing.

Preserving True AI Query Isolation: Most multi-agent systems rely on prompt engineering to "tell" the AI what not to leak. That's fragile. We had to architect strict query-level isolation in our database access layer so that a client's private AI assistant literally cannot query or inject embeddings from the freelancer's private workspace or conversation history, and vice versa.

Handling Scope Drift in Real-Time Conversations: Detecting verbal commitments during live video calls without interrupting natural conversation was a tough challenge. We had to tune our Groq/Gemini pipeline to extract actionable scope deltas into proposed ADR drafts, ensuring the AI never commits to anything on behalf of the user-every card requires two manual clicks to become active.

Deliverable Sandbox Limits: Running deep AST parsing, mutation tests, and visual browser snapshots against diverse student and enterprise repository archives without blowing past server memory limits required building streaming zip extractors and strict timeout wrappers.

Accomplishments that we're proud of

Building an impartial AI mediator, not just another chat app: The freelance world doesn't need another generic project tracker - it needs an end to scope creep and dispute anxiety. We successfully built a dual-AI mediation layer where each party has a private AI copilot that translates tense, vague conversations into clear, mutually beneficial agreement cards.

Zero-Downtime Multi-LLM Cascade: We didn't lock ourselves into a single AI provider. We engineered a resilient model mesh that routes between Google Gemini and DeepSeek-R1 , backed by deterministic rule-based fallbacks. If an LLM goes down, Nexum stays up.

A Truly Headless Telegram Command Center: We are especially proud of our omnichannel bridge. You can walk down the street, send a quick voice memo on Telegram, have Groq Whisper transcribe it in milliseconds, and watch it post into the web workspace with a SHA-256 cryptographic seal. The bot even supports @mentions and Swipe-to-Reply to juggle multiple live contracts without ever opening a laptop.

Bulletproof Financial & Contract Integrity: Every state change, amendment, and payment milestone is locked in an append-only audit trail. We treated money like real banking: integer-only minor units, zero floating-point math, and a strict human-in-the-loop principle where AI can draft, but only humans can approve escrow releases.

What we learned

AI is best as an emotional shock absorber, not a decision-maker: In freelance work, relationships sour over awkward money discussions or creeping deadlines. We learned that users don't want AI making binding decisions for them - they want AI to do the uncomfortable emotional labor of structuring trade-offs, calculating fair price deltas, and drafting polite compromise proposals.

Graceful degradation is non-negotiable in production AI: LLM APIs timeout, rate-limit, and experience latency spikes. Learning to design a multi-tiered fallback pipeline - from primary cloud LLMs down to 20-second timeout fallbacks and deterministic heuristic engines - was our biggest technical eye-opener. A product is only as intelligent as its ability to stay online when the AI hiccups.

Bilateral sync is tricky but magical: Bridging real-time web workspaces with the asynchronous nature of Telegram required careful event deduplication, echo-loop prevention, and conversational regex parsing. But the moment we saw a Telegram voice note seamlessly update a live web Kanban board, we knew the effort was worth every line of code.

What's next for Nexum

Autonomous CI/CD & Design Verification: Integrating GitHub webhooks and Figma APIs so Nexum AI can automatically run test suites, check linting, and inspect design tokens when a deliverable is submitted - giving clients objective quality metrics before approving escrow release.

Global Multi-Currency & On-Chain Escrow: Expanding beyond PayPal/Stripe to support self-custody on-chain smart contracts (USDC/USDT) for zero-fee cross-border payouts to international freelancers. Multi-Freelancer Agency Pods: Expanding our bilateral 1-on-1 mediation model into multi-party workspaces where an AI project manager coordinates deliverables across a designer, frontend dev, and backend engineer under one unified client budget.

WhatsApp & Slack Relays: Expanding our omnichannel command center to WhatsApp Business and Slack so teams can collaborate without leaving their preferred day-to-day messaging apps.