pk. Prashant Kumar
AI Architect · Author · Building POCs
Issue 01 · Vol. I Currently building — Avarna v4 · Stock Predictor · Multi-Agent — · 2026
·POC
02Product
Avarna Privacy Gateway · working POC

Privacy you can prove, not just promise.

A drop-in gateway between your apps and any LLM provider. Three detection layers in parallel, an encrypted entity graph, and a round-trip you can put on a chart. Status — working prototype; no company yet.

Avarna intercepts AI traffic before it leaves your perimeter — detecting PII with three layers in parallel, masking it, holding the cleartext-to-token map in an encrypted entity graph, and unmasking the model's response on the way back.

The prototype already supports full provider choice — Claude, GPT, open-weights — without the compliance headache. Self-hostable. Model-agnostic. With an auditable trail of every prompt and every byte that crossed the wire. It is a POC today, not a product yet.

Regex Microsoft Presidio LLM classifier Encrypted vault Audit trail Self-host

Pipeline · Detect → Mask → Entity Graph → Unmask

i.
Detect — in parallel
Three layers find PII before anything leaves.
Regex — high-precision patterns (emails, IDs, phones)
Presidio — Microsoft's NER-driven detector
LLM classifier — context-aware catches the other two miss
ii.
Mask
Each entity is replaced with a deterministic token.
iii.
Entity graph — encrypted vault
Token-to-cleartext map persists in an encrypted store. The LLM call then 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.
02Stock
Stock Predictor · early POC

A quantitative exploration.

Combining classical features with learned representations — a POC for testing whether modern ML pipelines can produce signal beyond the noise floor. Methodology and results land here as the experiment runs its course.

Status: in progress. Repo: [private until v0]. Watch this section for the writeup.

03Agents
Multi-Agent Platform · early POC

Substrate for agent collaboration.

A working prototype of how multiple agents can plan, share tools, and hand off context — with evals at every seam. Architecture diagram and first results will publish when v0 stabilizes.

Status: design + prototype. Repo: [private until v0]. Detailed writeup pending.

04Other
Other explorations

Small, sharp experiments.

Permissive-licensed tools, side experiments, and one-day hacks. Once the GitHub feed is wired up, the most recent will surface here automatically.

Status: placeholder. Drop GitHub username in memory/user_profile.md and the agent will pull repos in on the next weekly run.