💡 Inspiration

Every month, finance teams at fast-growing companies and multinational enterprises face a recurring nightmare: the month-end financial close. Accountants and finance controllers spend 14+ grueling days manually reconciling hundreds of PDF vendor invoices against raw bank debit records (NEFT, RTGS, IMPS wires) across dozens of disjointed, error-prone Excel spreadsheets.

A single missed TDS deduction under statutory regulations (such as India's Section 194Q or US SOX 404 controls) can lead to statutory fines, delayed audit sign-offs, and millions in unrecovered duplicate disbursements.

We asked ourselves:
Why are multi-million dollar finance departments still doing forensic accounting in legacy spreadsheets when high-frequency traders have Bloomberg Terminals?

We set out to create ReconX-AI: a zero-latency, institutional-grade financial reconciliation terminal that combines the cognitive horsepower of a 6-agent autonomous AI swarm with the clarity of Swiss graphic design, tactile acoustic ergonomics, and universal accessibility.

⚡ What It Does

ReconX-AI automates the entire counterparty reconciliation and statutory compliance lifecycle:

Autonomous Document Ingestion ([F2] Vault): Ingests vendor invoices (PDF, TXT, scanned images) and bank statement feeds (CSV), parsing metadata with dual PyMuPDF + Tesseract OCR and mathematical tax validation.
5-Axis Neural Vector Radar ([F4] Matcher): Matches invoices against bank debits across five weighted semantic and numerical dimensions, featuring an interactive What-If Sensitivity Lab.
X-Ray Laser Diff Workbench: Visualizes line-item correspondence using animated SVG Bezier laser beams, instantly bridging invoice items to bank UTR settlement lines.
Autonomous Anomaly Guard ([F5]): Flags subtle statutory discrepancies, including 0.1% Section 194Q TDS withholding variances, duplicate payment fingerprints, and unexpected wire fees.
ChromaDB Forensic RAG ([F6] Investigations): Automatically retrieves company accounting policies and statutory guidelines to compile evidence timelines and actionable accounting recommendations.
Immutable Cryptographic Audit Trail ([F10]): Every autonomous agent decision and human approval is sealed with SHA-256 cryptographic hashes and printable SOX-404 Statutory Compliance Certificates.
Inclusive Accessibility Suite: Certified WCAG 2.1 AAA with Okabe-Ito Colorblind-Safe Cyan, Swiss Daylight Paper, Dyslexia Hyperlegible Typography, and a Web Audio API Tactile Synthesizer.

🔬 The Math Behind the 5-Factor Vector Radar

Rather than relying on unconstrained generative language models that introduce dangerous numerical hallucinations, ReconX-AI formulates counterparty reconciliation as a multi-dimensional constrained optimization problem. The global match affinity $S_{\text{match}} \in [0, 1]$ between an invoice transaction $T_{\text{inv}}$ and a bank settlement transaction $T_{\text{bank}}$ is computed via a weighted affine vector distance:

$$S_{\text{match}} = w_v \cdot \Phi_{\text{vendor}} + w_a \cdot \Phi_{\text{amount}} + w_r \cdot \Phi_{\text{ref}} + w_d \cdot \Phi_{\text{date}} + w_n \cdot \Phi_{\text{narration}}$$

Subject to the convex simplex constraint:

$$\sum_{i \in {v, a, r, d, n}} w_i = 1.0, \quad \text{where } w_i \ge 0$$

Component Metric Scoring Functions:

Entity Vendor Affinity ($\Phi_{\text{vendor}}$, $w_v = 0.35$): Combines token-sort Levenshtein distance with character $n$-gram Jaccard similarity.
Amount Precision ($\Phi_{\text{amount}}$, $w_a = 0.35$): Governed by an inverse exponential decay function evaluated over absolute currency differential $\Delta = |A_{\text{inv}} - A_{\text{bank}}|$: $$\Phi_{\text{amount}} = \exp\left(-\frac{\Delta}{\tau_a}\right), \quad \tau_a = \max(1.0, 0.005 \cdot A_{\text{inv}})$$
Reference & UTR Verification ($\Phi_{\text{ref}}$, $w_r = 0.15$): Alphanumeric regularized longest common substring ratio over bank UTR identifiers.
Temporal Proximity ($\Phi_{\text{date}}$, $w_d = 0.10$): Linear decay over the settlement delta $\delta = |t_{\text{bank}} - t_{\text{inv}}| \text{ days}$: $$\Phi_{\text{date}} = \max\left(0, \, 1 - \frac{\delta}{7}\right)$$
Narration Coherence ($\Phi_{\text{narration}}$, $w_n = 0.05$): Cosine similarity between normalized 384-dimensional dense semantic sentence embeddings: $$\Phi_{\text{narration}} = \frac{\mathbf{e}{\text{inv}} \cdot \mathbf{e}{\text{bank}}}{|\mathbf{e}{\text{inv}}| |\mathbf{e}{\text{bank}}|}$$

🛠️ How We Built It

Frontend & Design System: Next.js 14 App Router, TypeScript, and Tailwind CSS. Built with high-density Bloomberg-inspired OLED palettes, SVG coordinate math for the real-time vector radar, and hardware capture-phase keyboard shortcuts (F1–F10 and Ctrl+K Omni-Search).
**Sens