Skip to content
View ganjikrishna's full-sized avatar

Block or report ganjikrishna

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ganjikrishna/README.md

Hi, I'm Vinay K Ganji Krishna

Information Technology Manager at Marcom Company, a media and marketing startup, where I have led a team of about 15 since October 2019. My background is 18+ years in enterprise IT across Horizon BCBS, Citigroup, IBM, and Infosys.

What I do

I run IT and put AI to work in production (Claude, ChatGPT, Gemini). My CRM depth is Siebel, Salesforce, and HubSpot. In public I build small, practical tools for marketing operations and CRM workflows, and I write up what they do and where they fall short.

Selected projects

CRMArena-Pro Gate: AI agents working in a CRM can be asked for data they should not reveal. I built and measured a pre-answer authorization gate (Gate Spec GS-2 v1.0) on Salesforce AI Research's CRMArena benchmark, B2B single turn tasks. Refusals of confidential requests rose from 0 of 75 to 42 of 75, while workflow scores held at 47 of 50 and policy scores at 10 of 100 in both runs. One request type, confidential company knowledge, still only reached 8%, and the repo states that plainly with the rest of its limits. Published November 13, 2026.

Signet: editing a language model using only plus and minus signs. On a sealed test, story degeneration in Qwen2.5-0.5B fell from 86.0% to 69.1% (p = 0.0002) after sign-only edits to the weights, with no training run. The repo reports the method, the sealed results, and where the approach hits its limits. November 2026.

RustCRM: an open CRM base written in Rust, with role based and field level permissions built in. It is in testing now and will be published here under Apache 2.0.

Earlier work

Fraction Compressor compresses the difference between a base model and its fine-tune into a patch file of 19 MB, against 988 MB for the full change, at about 0.31 bits per weight. It closed 23.2% of the fine-tune gap on 96 questions. It is lossy, and a distilled LoRA baseline beat it. Built for the Meta Global AI Developer Hackathon, October to November 2026, which ended without a win. That work is complete. The earlier version is at fraction-compressor.

Where I am heading

Quantum software and Quantum AI. I entered the Google Quantum AI circuit compilation competition on Kaggle on November 12, 2026 (team Vinay; no submission yet), and I learn in public with small projects, like the QRNG API, which serves randomness from real quantum measurement on IBM Quantum hardware.

Tools and expertise

CRM: Siebel, Salesforce, HubSpot

AI in production: Claude, ChatGPT, Gemini

Code: Python, Rust, JavaScript, REST APIs

Quantum: Qiskit, IBM Quantum

Elsewhere

LinkedIn: linkedin.com/in/vinay-gk

Based in Bethlehem, Georgia, near Atlanta.

Pinned Loading

  1. fraction-compressor-v0.4 fraction-compressor-v0.4 Public

    Python

  2. vibe-security-check vibe-security-check Public

    JavaScript

  3. signet-qwen25 signet-qwen25 Public

    Signet: fixing a language model by editing its weights with plus-or-minus steps only. Qwen2.5-0.5B, full sealed-test record.

    Python

  4. revenue-signal-copilot revenue-signal-copilot Public

    Python

  5. campaign-content-qa campaign-content-qa Public

    Python

  6. invoice-ops-assistant invoice-ops-assistant Public

    Python