I build systems
that think.

Voice and chat AI agents fail in ways nobody sees until a customer complains. I build VEE Legion, which finds those failures first with simulated callers and red-team attacks.

Concurrent calls

200

Transport errors

0

Platforms tested

5

// Work · VoiceraCX

The_ Legion

Independent QA for voice and chat AI agents

I defined the product and built the backend of VEE Legion. It tests AI agents the way customers do, with simulated callers and red-team attackers, then shows exactly where they break. Break the agent before your customers do.

legion.voiceracx.ai ↗
  1. 01 · The problem

    Which calls went wrong?

    An AI agent can handle thousands of calls a day. A few go wrong, and nobody knows which until a customer complains. VEE Legion finds them first.

    A wall of resolved calls with two flagged as failures · from the VEE Legion launch video
  2. 02 · The callers

    Callers who push back

    Scenario-driven callers are frustrated, confused, anxious, pushy, tech-savvy or on a noisy line. Red-team callers try prompt injection, jailbreaks and social engineering. All of them test the agent exactly as it is, on VEE, Twilio, Retell, Vapi or Synthflow.

    Twelve caller personas shown as cards · from the VEE Legion launch video
  3. 03 · The rules

    Define a good call in plain words

    Pass or fail can be a sentence, such as “verify the caller’s identity before changing account details”, judged by an LLM next to automated checks for WER/CER, latency, tool calls, PII and keywords. A failed judge run never counts against the agent.

    The create-a-test wizard choosing callers and pass/fail rules · from the VEE Legion launch video
  4. 04 · At scale

    Hundreds of calls at once

    Slot-based orchestration in Go ramps load up instead of firing everything at once, so the test measures the agent and not the test rig.

    A test run executing many calls in parallel · from the VEE Legion launch video

    200

    concurrent calls verified

    0

    transport errors

  5. 05 · The why

    Know exactly why it failed

    Every call is scored and every failure is shown turn by turn. Real production calls can be imported and graded with the same checks, so tests and reality sit on one scale.

    A transcript with the failing agent turn stamped FAIL · from the VEE Legion launch video
  6. 06 · Try it

    Test an agent

    Pick a caller and run the test. These conversations are illustrative, written for this demo, but the checks work the way Legion’s do.

    Goal · Change my billing address, quickly.

    Verify identity before account changesNever promise fee waiversStay on task

    Press “Run test” to start the call.

    Illustrative conversations written for this demo, not real call data.

// Manifest_v5.0

The_ Architecture

Three layers of a voice agent that I have built and measured: how it talks, what it knows, and the models behind it.

The generic path of a voice agent. The solid stages are parts I built, across VoiceraCX, Alesa AI and the Orpheus TTS project.

Voice systems

3s → 1.2s

Real-Time Voice Infrastructure

Built LiveKit voice pipelines with streaming inference, connection pooling, and smart caching, reducing response time from 3 seconds to 1.2 seconds.

Retrieval

−90%

Hybrid Vector Search

Migrated Chroma to Qdrant with hybrid dense-sparse search, reducing query latency by 90% for voice-agent workloads.

Models

10k+ terms

Fine-Tuning & Evaluation

Fine-tuned Llama-3.1-8B with PEFT on a 10,000+ term dream-interpretation dataset and trained Orpheus TTS to generate SNAC audio tokens.

// Work · Education

The_ Journey

Drawn like a commit graph: my work on the main line, and my degree running alongside it as a branch that merged in July 2026.

  1. HEAD → main⑂ merged pict · Jul 2026

    June 2026 — Present · Pune

    VoiceraCX

    AI R&D Engineer · VEE Legion

    Building VEE Legion, an independent QA platform that tests voice and chat AI agents with simulated callers, automated scoring and red-team attacks.

    • Defined the product scope and engineered VEE Legion, a production platform that tests unmodified voice and chat agents with scenario-driven callers across VEE, Twilio, Retell, Vapi and Synthflow.
    • Built Go-based concurrency ramping and slot-based orchestration over Twilio and LiveKit; validated 200 concurrent calls with zero transport errors through controlled load testing.
    • Developed automated evaluation across WER/CER, latency, tool calls, PII, scenario checks and LLM-as-a-judge, with safeguards against failed judge executions skewing pass rates. Added red-team callers, turn-level failure analysis and real-call import.
    200calls verified0transport errors5platforms testedSee VEE Legion
  2. June 2025 — June 2026 · Pune

    VoiceraCXArtificial Intelligence Intern

    Real-time voice pipelines and scalable vector retrieval for voice-agent products.

    3s → 1.2sresponse time−90%query latency−60%API calls
  3. Feb 2025 — June 2025 · Remote

    DAOStreetSoftware Development Intern

    Svelte web application features in a production workflow.

  4. Nov 2024 — Mar 2025 · Remote

    Alesa AI Ltd, UKAI/ML Intern

    Tangent Mind, a tarot, horoscope and dream-interpretation platform.

    10k+datasetLlama 3.1 8Bmodel
  5. ⑂ pict/b.e. · ran alongside everything above

    Nov 2022 — Jul 2026 · Pune

    Pune Institute of Computer Technology

    Bachelor of Engineering in Computer Engineering · Honors in Data Science

Manifest_v7.0

The_ Forge

Seven projects, from voice models to company memory. Open any tile for its architecture and design decisions, or ask Wisp to explain it.

Manifest_v7.0

The_ Forge

Seven projects, from voice models to company memory. Open any tile for its architecture and design decisions, or ask Wisp to explain it.

Explore the projects

FalkorDB

October 2026

Weave

A company brain that knows who's allowed to know what

FalkorDBCypherPythonFastAPIPostgres + pgvectorMCP+3

7

Connectors

13

Graph node types

12

Agent tools

Design questions it answers

  • ›Why a graph instead of vector search over documents?
  • ›Why enforce permissions inside the Cypher query?
  • ›How do you keep a claim only if it has visible evidence when FalkorDB lacks EXISTS subqueries?
  • ›Why supersede facts instead of overwriting them?
  • ›How does impact analysis avoid leaking private threads?
  • ›Why do agents call fixed tools instead of writing Cypher?
  • ›Who is the audience in a shared Slack channel?
GitHub ↗

Generated by the model

Featured

Llama 3.2 3B

March 2026

Orpheus TTS

A language model where the 'language' is sound

Llama 3.2 3BLoRAUnslothSNAC Codec+4

19min

1 epoch, 299 steps, single GPU

GitHub ↗
React FrontendFastAPI BackendPostgreSQLQdrantCrewAI AgentsAI CalendarExam Generator
Featured

FastAPI

July 2025 – June 2026

Smart Pathshala

Offline-first AI school ecosystem for Tier-2 and Tier-3 India

FastAPIReactPostgreSQLQdrant+1

4

Admin, Teacher, Parent, Student

lockPrivate
BloomLog screenshots

Next.js 16

May 2026

BloomLog

A gentle daily companion for mood, water, meals and tiny wins

Next.js 16React 19Tailwind CSS 4Framer Motion+4

60sec

Mood, water, meals, sleep and tiny quests in about a minute

Live ↗lockPrivate

LLM

April 2025

ShetNiyojan

Intelligent agricultural planning from seed to sale

LLMFlaskMongoDB

1st

Winner at Devclash 2025, DY Patil

GitHub ↗
WarCast screenshots

Flask

2025

WarCast

AI-based defense news aggregator with sentiment analysis and summaries

FlaskPythonDistilBERTBART+2

10+

Global defense publishers aggregated in real time

GitHub ↗
Legify screenshots

BERTSum

March 2025

Legify

AI-powered legal document simplification and Q&A

BERTSumFAISSDjangoTTS/STT

1st

Winner at Synapse 2.0, CCOEW

GitHub ↗

// System_Status: Operational

The_ Matrix

The tools I build with, grouped by what they are for.

  • Python
  • PyTorch
  • Go
  • C++
  • TypeScript
  • React
  • Next.js
  • Svelte
  • Tailwind CSS
  • FastAPI
  • PostgreSQL
  • MongoDB
  • MySQL
  • Docker
  • Kubernetes
  • Qdrant
  • LiveKit
  • Twilio
  • LangChain
  • CrewAI
  • Pipecat
  • OpenTelemetry
  • Jaeger
  • Lightning AI
  • AI / ML: PyTorch, LangChain, LLMs, RAG, LoRA / PEFT, Unsloth, CrewAI, SNAC Codec
  • Voice & Inference: FastAPI, Pipecat, LiveKit, Twilio, llama.cpp, GGUF
  • Evaluation & Testing: LLM-as-a-judge, WER/CER, Latency benchmarking, Red-teaming, PII checks, Load testing, Twilio, LiveKit
  • Data & Infrastructure: Qdrant, PostgreSQL, MySQL, MongoDB, Docker, Kubernetes, Helm, OpenTelemetry, Jaeger, KEDA, ChromaDB, FAISS, Lightning AI
  • Languages: Python, C++, Go
  • Frontend: React, Next.js, Svelte, TypeScript, Tailwind, Framer Motion

// Recognition

Achievements_

1st prize

Devclash 2025

24-hour hackathon at DY Patil Pimpri

1st prize

Synapse 2.0

AI hackathon at MKSSS CCOEW

1st prize

L&T NeuroHack

24-hour hackathon at COEP Mindspark 24

99.35 percentile

MHT-CET 2022

Merit rank 971 among 400,000+ candidates

// Contact

Let's build something.

Tell me about the role, the project or the idea. The quickest way is email, and the form below opens your mail app with everything filled in.

This opens your email app. Nothing is sent from this page.