AI builder · Manager @Accenture

Vedavyas Vayalpadu

Enterprise AI from pilot to production, built on years of running AIX, IBM Power and SAP HA.

See the tools Contact
Vedavyas Vayalpadu

Production infrastructure roots.
AI tools, built every day.

IBM AIXIBM PowerSLES for SAP HAPacemakerAzureClaude CodeEvalsRAGAgent ops
6tools I built and use
253tests behind ADDA on PyPI
211findings in a 40-repo AI study
39diagnostic parsers, 3 platforms

Six tools I built.
Each one started as a problem I had.

AI quality.
Catch the wrong answer before users do.

Two tools for the failure that matters most in AI: the confident, plausible, wrong answer.

The method ADDA on PyPI

ADDA

Shipped · PyPI

Stops a commit when the code changed and its docs didn't, so AI agents never read stale context.

  • v0.5.4
  • 253 tests
  • Offline
  • MIT
PyPI GitHub Site

AI-AppLens

Research

Reads an AI codebase without running it and scores whether its drift could even be detected.

  • 3-drift taxonomy
  • 40 repos
  • 67 high / critical
Private repoThe method behind it

Infrastructure.
The diagnostics I wanted on call.

Support bundles and IBM Redbooks, turned into answers an engineer can act on.

LinuxDiag

Live beta

Turns a 300 MB sosreport, supportconfig or AIX snap into a health score and ranked findings.

  • RHEL · SUSE · AIX
  • 39 parsers
  • No outbound HTTP
Private repo

IBM-BOB-AIX

Learning lab

My Power and AIX lab: a local RAG assistant over 15 IBM Redbooks, with measured retrieval.

  • 19 / 20 retrieval eval
  • Qwen or Haiku
Private repo

Everyday AI.
Tools I use every day.

Speech in and out in five languages, and study notes from every call and lecture.

VoiceForgeAI

Working

Ask a GitHub repo a question out loud and hear two grounded sentences back.

  • EN · HI · KN · TA · TE
  • ~530 commits
Private repo

ShadowShadow

Daily use

Transcribes my calls and lectures on the laptop and turns them into study notes. No audio leaves the machine.

  • faster-whisper
  • Local
  • 11 test suites
Private repo

One lens for AI systems.
Three kinds of drift.

01

Model drift.

Same prompt, different answer after a silent provider change.

Caught with a frozen golden set.
02

Behavioral drift.

Quality decays inside one long run as the context fills.

Caught with a per-step quality slope.
03

Agentic drift.

The agent's tool choices shift over time.

Caught against a tool-call baseline.

The dangerous defect is not the one that breaks. It is the confident, plausible, wrong answer.

Writing.
What I post on LinkedIn.

All posts

Four tasks that look right and are wrong

The worst answer is not the one that is missing. It is the one that was invented.

Read on LinkedIn

RICE and CRAFT, two frameworks I use daily

The test for an enforcement rule is whether you can fail it. Most instructions we give AI are wishes, not rules.

Read on LinkedIn

The Forward Deployed Engineer

I used to think my value was in knowing the syntax. Now I think it is in knowing which problem is actually the problem.

Read on LinkedIn

Claude Code's Ultracode mode

A big signal for where software delivery is heading.

Read on LinkedIn

Learning in public.

  1. FDE Experience Day, Pronative.aiDiscovery to hardening to runbook, hands-on.
  2. NASSCOM AI-Code SarathiForward-deployed engineering, RICE and CRAFT.
  3. NVIDIA + OpenACC hackathon finalsBengaluru, participation certificate.
  4. NVIDIA workshopNemotron, agentic workflows, NeMo Agent Toolkit.
  5. Google Build with AI bootcampCompared Antigravity with my Claude Code workflow.

Let's talk about AI that holds up in production.

Message me on LinkedIn GitHub