ComplyVision is an AI-powered live compliance monitoring and audit-intelligence platform built for Indian manufacturing environments. It is designed to serve two connected audiences in the same workflow: industry compliance and internal audit teams operating on the shop floor, and Chartered Accountants in practice who rely on evidence from that shop floor to form an audit opinion.

The platform is built around a single positioning idea: move compliance from periodic, post-facto checking to continuous, evidence-ready assurance.

Most statutory and operational compliance in Indian manufacturing today is verified through point-in-time checks — an annual factory inspection, a quarterly internal audit walkthrough, a monthly statutory filing reconciled after the fact. ComplyVision replaces episodic checking with a continuously running monitoring and evidence layer that a CA can sample from at any point during an engagement, rather than relying solely on a single observation date or management representation.

Three Pillars

1. Live operational compliance monitoring — AI-based vision detection on the shop floor for PPE compliance, restricted zone entry, unsafe stacking, and fire/smoke risk.

2. Document verification and statutory evidence review — upload of a Professional Tax (PT) challan; the system extracts, classifies, and cross-checks it against statutory compliance master data.

3. Audit-ready exception reporting and escalation — every flagged exception is converted into a structured, evidence-backed action item formatted for a CA's working paper.

2. Problem Addressed

Statutory and operational compliance verification in manufacturing today is largely manual, periodic, and disconnected from the evidence that supports it. This creates four specific, recurring problems for both industry compliance teams and the CAs who audit them.

Problem

Description

Episodic verification

PPE compliance, restricted zone access, and shop-floor safety controls are typically checked through periodic physical walkthroughs — once a quarter or once during an annual audit. Violations occurring between checks go entirely undetected and undocumented.

Manual document cross-checking

Statutory filings such as Professional Tax challans are verified manually against salary registers — a repetitive, error-prone task performed every month for every client, with no structured audit trail of the verification performed.

Disconnected evidence

Even when a violation or exception is identified, the evidence — a photograph, a note, a verbal observation — rarely makes it into a structured, working-paper-ready format. CAs spend disproportionate time converting raw observations into auditable documentation.

Sampling-based assurance

Because monitoring is not continuous, audit conclusions rest on a small sample of observations from a single point in time. This limits the strength of the assurance a CA can reasonably provide on operational and statutory compliance matters.

ComplyVision is built to directly address these four gaps without displacing the CA's professional judgement — it strengthens the evidence base the CA works from, rather than replacing the CA's conclusion.

3. Solution and Process Flow

ComplyVision operates as a continuous loop across three stages: detection, verification, and escalation. Each stage produces structured output that feeds the next, ending in a CA-reviewable audit action.

Stage 1 — Live Operational Monitoring

4. Camera feeds are processed through a YOLOv8n object-detection model, fine-tuned on a PPE dataset (hard hat, safety vest, gloves) via Roboflow weights.

5. Detected violations (missing PPE, restricted zone entry, overstacking, fire/smoke indicators) generate a timestamped incident with camera ID, location, severity, and AI confidence score.

6. Employee identification, where available, is captured separately through an opt-in QR badge scan (pyzbar) — the vision model itself does not perform facial recognition or identify individuals, by design.

Stage 2 — Document Verification

7. A compliance document (e.g. a Professional Tax challan) is uploaded against an open compliance task.

8. The system extracts key fields — challan type, state, entity, period, amount, payment date — and validates them against a statutory compliance master and the client's salary register.

9. A deterministic rule engine issues one of four verdicts: valid, invalid (amount mismatch, with expected vs detected figures shown), unrelated (wrong document type, e.g. PF submitted instead of PT), or incomplete (blank or unreadable document).

10. A human reviewer (CA or compliance officer) can override any verdict with documented remarks; the original AI verdict is preserved alongside the override for audit trail purposes.

Stage 3 — Escalation and Audit Action

11. Every open incident or failed document carries a clear flag that CA review and sign-off is required before any action is finalised.

12. A single action converts the exception into a CA-style working-paper observation: area, control, observation, risk and impact, management response field, responsible party, priority, and status.

13. The same action generates a draft management email and a structured JSON payload for workflow automation (e.g. n8n, Slack, Teams) — giving the CA both the working paper output and a ready escalation artefact in one step.

Architectural principle

Detection and compliance logic are deliberately decoupled. The vision model determines what is happening in a frame; a separate, explainable rule engine determines whether that constitutes a compliance exception under a specific statutory or procedural rule. This separation means the platform can be extended to new domains — construction safety, food hygiene, pharma GMP, cash-handling controls — by retraining the detection model and reconfiguring the rule set, without rebuilding the underlying platform.

4. Key Features and Controls

Feature

What It Does

Live monitoring dashboard

Real-time camera grid with AI-detected compliance status per feed, severity-tagged alert stream, and a compliance health score for the site.

PPE & zone-entry detection

YOLOv8n-based object detection for hard hat, safety vest, glove compliance, and restricted-zone entry, with confidence-scored bounding boxes.

PT challan verification

Document upload and field extraction, validated against statutory master data with a deterministic rule engine; production-mode OCR + vision-LLM extraction available via a feature flag.

Generate Audit Action

One-click conversion of any incident into a structured CA working-paper observation, a management email draft, and a webhook-ready JSON payload.

Human override / reviewer sign-off

Every AI verdict — document classification or detection-based incident — can be overridden by a human reviewer with documented remarks; original AI output is retained for audit trail.

Role-based access control (RBAC)

Routes and actions are gated by user role, separating compliance officer, auditor, and admin access levels.

Privacy-by-design architecture

Vision inference is designed to run on-device; only compliance event metadata is retained, not raw video or biometric facial data. No employee facial recognition is performed.

Statutory compliance calendar

Tracks PF, ESI, PT, TDS, and other recurring statutory deadlines with task-level status, ownership, and document evidence linkage.

Built-in Controls

• AI risk indicators are explicitly labelled as requiring CA professional judgement before being treated as a final conclusion — never presented as an autonomous finding.

• A hybrid validation model (deterministic rule engine plus optional AI extraction) ensures a model hallucination cannot silently pass an incorrect statutory document — the rule engine must independently agree.

• Document uploads are not persisted across sessions in the current build, reducing the retention surface for sensitive client data.

• Every override and escalation action is logged, preserving a clear chain from AI flag to human disposition to final action — directly supporting SA 230 documentation standards.

5. Impact

ComplyVision's impact is best understood across three stakeholders, each of whom currently absorbs a different cost from the problems described in Section 2.



Stakeholder

Impact

Chartered Accountant in practice

Moves from a single-point-in-time observation to a continuous evidence base. The PT challan exception report directly supports Section 43B disallowance analysis. The audit action output is pre-formatted for working papers, reducing documentation time per exception. Creates a basis for a recurring compliance-monitoring retainer as a new service line.

Industry compliance / internal audit team

Real-time visibility into PPE and zone-access violations instead of discovering them at the next scheduled walkthrough. Faster exception-to-resolution cycle through structured escalation rather than informal verbal reporting.

The organisation as a whole

A documented, continuous compliance trail strengthens the organisation's position in the event of a regulatory inspection, insurance claim, or workplace incident investigation — independent of whether any single exception was material.

Qualitative outcome

The core shift is from “we sampled and found nothing” to “we monitored continuously and these are the exceptions that were reviewed and resolved.” That is a materially stronger evidentiary basis for the CA's audit opinion, and a materially stronger value proposition for the engagement.

6. Tech Stack

Layer

Technology

Frontend

Next.js 14 (App Router), TypeScript, Tailwind CSS, shadcn/ui, Framer Motion, Recharts, Zustand

Backend

FastAPI, SQLAlchemy ORM, SQLite (MVP data store)

Computer vision

YOLOv8n (Ultralytics) with Roboflow-sourced PPE detection weights; runs as a decoupled worker process

Document AI (production path)

OCR + PyMuPDF for extraction; OpenAI vision-capable model for unstructured document parsing — flag-gated, off by default in demo mode

Identity capture

QR badge scanning (pyzbar) — opt-in, non-biometric employee identification

Architecture pattern

Hybrid AI + deterministic rule engine for all compliance verdicts; human override layer on every AI-generated conclusion

Design choices worth noting

• SQLite and a single FastAPI process are appropriate for the current MVP stage; a production, multi-location deployment would move to PostgreSQL with horizontally scalable, independently deployed CV worker processes.

• The AI document extraction path exists in the codebase but is disabled by default — the live demonstration uses a deterministic rule-engine path for reliability, with the AI extraction flag (USE_AI_DOC_VERIFIER) available to enable on request.

• No facial recognition or biometric identification model is used anywhere in the architecture; all vision inference is limited to object/PPE detection.

7. Future Roadmap

Horizon

Planned Capability

Near term

• Multi-state Professional Tax challan support (currently Maharashtra format in production extraction).

• Tally plugin integration so a flagged compliance exception automatically tags the relevant expense ledger entry.

• Expanded statutory deadline coverage across GST, TDS, and state-specific labour filings.

Medium term

• Voice-based external confirmation workflow under SA 505 — automated outbound call or WhatsApp confirmation request to a client contact, with the response captured as an auditable record.

• Partner-review dashboards aggregating compliance health across multiple client engagements.

• CA-firm client portal for direct client access to compliance status and document upload.

Longer term

• Domain extension via model retraining and rule reconfiguration — construction site safety, food/pharma hygiene (FSSAI, Schedule M), cash-handling controls for NBFC/banking correspondent audits.

• Aggregate, anonymised compliance pattern analysis across client engagements to support CA risk-scoring and engagement scoping decisions.

• Migration to a horizontally scalable, multi-location production architecture (PostgreSQL, independently scaled CV worker fleet).

8. Conclusion

ComplyVision addresses a structural gap in how operational and statutory compliance is currently verified in Indian manufacturing — not by replacing the Chartered Accountant's judgement, but by replacing the episodic, manually-assembled evidence the CA has historically had to work with.

The platform's architecture is deliberately built around three disciplines that matter as much as the underlying AI capability: a clear separation between detection (what the model sees) and compliance logic (what counts as a violation under a specific rule), a human review step on every AI-generated conclusion, and an explicit data-minimisation posture given the sensitivity of camera-based monitoring of people at work.

In its current form, ComplyVision is a working MVP demonstrating this architecture end-to-end across live monitoring, document verification, and audit-action generation. The roadmap extends the same architecture — not a different one — into adjacent compliance domains and a production-scale deployment, while preserving the core principle that the system generates risk indicators and evidence, never final legal or audit conclusions.