pk.Prashant Kumar
AI Architect · Author · Building POCs
Issue 01 · Vol. I Currently building — Avarna v4 · Stock Predictor · Secure Multi-Agent — · 2026
← All POCs
01Privacy
Working POC · v4 · benchmarks live

Avarna Privacy Gateway.

A drop-in gateway between your apps and any LLM provider. Three detection layers running in parallel, an encrypted entity graph, and a round-trip you can put on a chart. This page is the technical deep dive.

·Watch
3-minute walkthrough

See it in motion.

Drop the demo at /assets/media/avarna-demo.mp4
·How
Architecture

Detect → Mask → Entity graph → Unmask.

Three detection layers run in parallel — not the LLM alone. Each entity gets a deterministic token. The cleartext-to-token map is held in an encrypted vault. The masked prompt goes to the model. The response is rehydrated server-side on the way back.

The result: full provider choice — Claude, GPT, open-weights — without giving up PII control, compliance posture, or audit trail. Self-hostable. Model-agnostic.

Regex Microsoft Presidio LLM classifier Encrypted vault Audit trail Self-host
Request a walk-through ↗ See the deck ↓

Pipeline · in parallel where possible

i.
Detect — three layers in parallel
High-precision regex catches obvious PII fast. Presidio NER finds named entities. An LLM classifier picks up the contextual cases the first two miss.
Regex — emails, phones, IDs, card numbers
Presidio — Microsoft's NER-driven detector
LLM — context-aware classification
ii.
Mask
Each entity is replaced with a deterministic token so the model still sees structure, just no PII.
iii.
Entity graph — encrypted vault
Token-to-cleartext map persists in an encrypted store. The LLM call happens against masked prompts only.
iv.
Unmask
The model's response is rehydrated server-side using the vault — never client-side.
RTX 4070 · workstation
~0.55s round-trip
RTX 4080 · workstation
~0.40s round-trip
RTX 4090 · workstation
~0.25s round-trip
A100 / H100 · data-center
~0.15s round-trip
AWS g5
~0.35s round-trip
Azure NC-A100
~0.15s round-trip
GCP g2
~0.35s round-trip
Detection accuracy
97% PII recall, production
Round-trip = detect (parallel) + mask + entity-graph write + LLM call + unmask. Overhead vs. classification alone: ~30 ms.
·Deck
The pitch & the architecture

The deck.

Same content as above, in slide form. Open in a new tab for the full-screen experience.

If the embed is blank: drop the deck PDF at assets/media/avarna-deck.pdf — or replace the iframe with a Google Slides / SlideShare embed.

·Next
Where it's going

What's next.

Avarna is a POC today, not a product. The benchmarks above are real. The architecture is real. What's not yet built: a managed offering, a customer-friendly install path, the policy-engine UX that lets compliance teams write rules without reading code.

If you're a team with a privacy-sensitive AI use case and you'd be willing to run an early build, I want to talk. Same if you're an investor or potential collaborator who sees the same gap I do. Email's at the bottom of every page.