Product Case Study · Sovereign AI & Data Governance
Keeping data out of shadow AI — an offline PII firewall.
From a hackathon idea to Kalpita Sentinel: a fully developed, on-premise AI data firewall that detects, masks and grades personally identifiable information before it ever reaches ChatGPT, Copilot, Claude or any cloud model.
≥ 0.91
entity-level F1
≤ 3 s
p95 latency per chunk
100%
offline, provably
$0
per scan, per seat
Overview
Six things Sentinel guarantees about your data
Data never leaves the building.
Sentinel detects, masks, classifies and reports sensitive data on-device before it reaches any AI tool.
Hybrid by design.
A deterministic detector (Presidio + Piiranha + custom Indian-PII recognizers) pairs with an offline reasoning model (Phi-4-mini via Ollama) that writes a Red/Amber/Green verdict and a cited rationale.
Built for the regulation.
Engineered for India’s DPDP Act 2023 — ₹250 crore per violation, full enforcement 13 May 2027 — and mapped to GDPR, HIPAA, PCI-DSS, CCPA/CPRA, SOC 2, ISO 27001/27701, NIST 800-122 and FERPA.
Laptop-class, zero cloud cost.
Runs on a developer laptop with no GPU and no internet, at $0 per scan, scaling to an air-gapped enterprise server.
Proven outcomes.
≥ 0.91 entity-level F1, ≤ 3 s p95 latency, and a masked file, verdict and audit log in under ten seconds — measured offline.
Fully developed.
Every modality ships — text, documents, spreadsheets, PDFs, images, audio and video — with reversible tokenization, multilingual routing and a no-code recognizer builder.
Summary
The leak happens at the prompt. Sentinel stops it there.
Kalpita Sentinel is an offline, on-premise PII governance scanner — an AI data firewall that detects, masks and grades sensitive data before it ever leaves a user’s device. Built as a Kalpita internal innovation, it closes the fastest-growing enterprise data-leak channel of 2026: employees pasting customer, financial and health data into generative-AI tools. Industry data shows 18% of employees paste corporate data into GenAI tools, and over half of those pastes contain sensitive content (LayerX, 2025), while India’s DPDP Act now carries penalties of ₹250 crore per violation. Sentinel runs entirely on a laptop — no cloud, no internet, no per-scan cost — and returns a masked file, a Red/Amber/Green verdict and an audit log in under ten seconds. It is the pre-flight check every enterprise needs before a single prompt leaves its network.
Product profile
At a glance
This is a Kalpita internal-innovation story — Kalpita is both the builder and the first customer. The same teams that ship enterprise AI for regulated clients hit the shadow-AI problem first-hand, then built Sentinel to solve it.
- Product
- Kalpita Sentinel
- Industry
- AI-first technology — custom software, cloud and AI/ML engineering — with deep specialization in InsurTech, FinTech and HealthTech, the regulated industries where PII governance is non-negotiable.
- Size & profile
- 130+ AI/ML specialists, 8+ years of delivery, offices in Bengaluru and Austin, SOC 2 (AICPA) certified, ISO certified, Great Place to Work, Clutch 4.7.
- The challenge Kalpita set itself
- give regulated clients — and its own engineers — the productivity of generative AI without leaking regulated data to third-party models.
- Origin
- From a hackathon idea to a fully developed on-premise product
Tools & technologies
The stack behind the firewall
- L1 — detection
- Microsoft Presidio (MIT) · Piiranha v1 — 17 PII types, ~99% accuracy (MIT) · ONNX Runtime int8 (MIT)
- L2 — reasoning
- Phi-4-mini 3.8B, strict JSON (MIT) · Llama 3.2 3B fallback (Llama Community) · Qwen 2.5 3B multilingual (Apache 2.0) · OpenAI Privacy Filter (Apache 2.0) · Ollama runtime (MIT)
- Vision, speech & docs
- PaddleOCR — 100+ languages (Apache 2.0) · Qwen2.5-VL image-as-PII (Apache 2.0) · whisper.cpp + ffmpeg (MIT/LGPL)
- Parsing & validation
- pdfplumber · python-docx · pandas · lxml (MIT/BSD) · Pydantic / Ajv schema validation (MIT)
The challenge
The problem we solved
Industry data shows 18% of employees paste corporate data into GenAI tools — and under India’s DPDP Act 2023 each violation carries a ₹250 crore ceiling.
Generative-AI adoption has outrun governance. The leak is happening today, the regulator is awake, and most enterprises are still unprotected.
Shadow AI
industry data shows 77% of enterprise LLM use is ChatGPT, 18% of employees paste corporate data into GenAI tools, and 34.8% of ChatGPT inputs now carry sensitive data, up from 11% in 2023 (LayerX 2025; IBM 2025).
Regex gaps
pattern matching catches a PAN or an Aadhaar number, but not a surname in a sentence or a quasi-identifier combination — exactly the leaks an auditor cares about.
Cloud risk
every cloud LLM call is metered, and every prompt with PII is a data-residency and breach-notification risk.
₹250 crore
dPDP Act 2023 (₹250 crore per violation), GDPR, HIPAA and PCI-DSS all apply the moment sensitive data reaches a third-party model; full DPDP enforcement begins 13 May 2027 with no grace period.
Fragmented
security and compliance leads needed one tool — drag-and-drop simple, yet auditable enough for an ISO 27701 review.
The solution
A hybrid detector that never touches the network.
Sentinel turns an uncontrolled AI habit into a governed AI workflow — deployed inside the environment and configured to each customer’s industry, languages, data sources and compliance regime.
Offline by design
No telemetry, no internet, demonstrable with an airplane-mode toggle. The entire stack is open-licensed end to end — no GPL, no commercial-use surprises.
Layer 1 — deterministic detection
Microsoft Presidio and the Piiranha model, plus custom Indian-PII recognizers, catch structured identifiers in under a millisecond per chunk, and never hallucinate.
Layer 2 — contextual reasoning
A small offline model (Phi-4-mini via Ollama) reads context, proposes masks, writes a one-sentence rationale, and assigns the Red/Amber/Green verdict — returning strict, schema-valid JSON every time.
Reversible tokenization
Each entity becomes a stable token (Ms. Iyer → [PERSON_e3a2]) that can be swapped back inside the customer’s environment, so masked text can safely reach any cloud model, then be rehydrated locally.
A compliance citation on every entity
Every flagged item carries the clause it triggers — ’DPDP §2(t)’, ’GDPR Art 4(1)’, ’HIPAA 164.514(b)(2)’ — so the audit log is the evidence a regulator asks for.
Deploy where your data lives
Choose a one-click Windows desktop app for individuals, a Docker-packaged service with a browser-based interface for teams, or an air-gapped server deployment for enterprises and government organizations.
Inside the product
What ships in Sentinel
Every modality an enterprise actually handles — text, documents, spreadsheets, PDFs, images, audio and video — with an audit row behind each scan.
Every data source, one engine
text, CSV/Excel, JSON/XML, Word; text and scanned PDFs (OCR); images as text and as PII (faces, ID cards, signatures); audio and video via speech-to-text and frame extraction.
India-first PII coverage
Aadhaar, PAN, GST, IFSC, voter ID and Indian mobile formats, alongside multilingual OCR most global tools omit.
Reversible tokenization, caching, and streaming
safe round-trips to cloud models, a 24-hour cache, and streaming output that halves perceived latency.
No-code custom-recognizer builder
a compliance lead adds a new pattern through a regex builder in the UI, with no engineering ticket.
One-click audit log
every scan emits a timestamped Excel/CSV row: file, entities, verdict, model version, prompt version.
Multilingual routing and fairness guardrails
non-English chunks route to a multilingual model; precision and recall are tracked by name origin to catch bias.
Graceful degradation
if the reasoning model times out, Sentinel returns the deterministic result with a clear warning and a conservative verdict.
Guardrail modes for the whole AI stack
a browser extension intercepts pastes into ChatGPT; an IDE plugin scans prompts; a RAG pre-ingestion mode scrubs documents before embedding; an egress guardrail screens agent and MCP-server outputs.
Inside Kalpita Sentinel
Results
What Sentinel delivers, measured
Every figure below was measured offline, on laptop-class hardware, and verified against a separate evaluation set.
Detection quality
≥ 0.91 F1
≥ 0.91 entity-level F1 on a 100-chunk evaluation set spanning structured, contextual, multilingual, adversarial, and clean text — verified by a separate output-verification team
Latency
≤ 3 s p95
≤ 3 s p95 latency per 512-token chunk on laptop CPU, with the deterministic layer responding in under a millisecond
Cost
$0 per scan
against a cloud equivalent, a 5,000-employee deployment avoids roughly $108,000 a year in inference fees — before counting breach liability removed.
Network
100% offline
an airplane-mode toggle and zero telemetry let a regulator watch the network stay dark while a document is scanned.
Exposure avoided
₹250 crore
one masked KYC transcript keeps Aadhaar, PAN and account data out of a cloud model and out of DPDP scope.
Audit trail
Under 10 sec
a masked file, a compliance verdict and a logged evidence row in under ten seconds.
Conclusion
Sovereign AI, proved with an airplane-mode toggle.
Kalpita Sentinel proves a thesis Kalpita lives by - the safest AI is the one that never sees your secrets in the first place. By keeping detection, masking, and grading entirely on-device, Kalpita gives regulated enterprises the productivity of generative AI without the data-leakage liability — at zero per-scan cost, on hardware they already own. The journey from a single hackathon weekend to a fully developed, multi-modal platform is the clearest expression of how Kalpita works: practical innovation, shipped with discipline, governed by evidence — and a preview of the pipeline, as more AI products and solutions graduate from Kalpita’s continuous R&D.
FAQ
Frequently asked questions
See it live
See it run offline.
Book a 30-minute sovereign-AI readiness call and watch a document get masked, graded and logged with the network switched off.