01RAJVEER VADNAL / AGENTIC ENGINEER

Open to AI engineering roles & freelance

I build systems
that can prove
themselves.

I build tools that check whether coding agents actually did the job. Proof of Work re-runs tests and records verdicts; Tessera helps teams plan and improve tests. I also co-built a WhatsApp assistant for Solapur Municipal Corporation.

Explore my work Experience

Based in Solapur / Hyderabad, India
Building in public. Learning by shipping.

A handwritten problem, mechanical computing module, verification screen, and checked delivery crate connected on a workbench. — FIG. 01 — THE WORKING LOOP
A useful system starts with a real human problem.

HUMAN INTENT VERIFIABLE OUTCOMESDownload résumé ↓LOCAL TIME IST

02 / THE METHOD

The proof
is in the loop.

Ideas mean nothing without evidence. Here’s how a real problem becomes a trusted outcome.

Follow the architecture ↗

01 / INTENT
“Build a tool
that can test
itself.”A human problem
02 / CONTEXT
- Codebase
- Documentation
- Constraints
- Real-world usage
- Previous failures
03 / ACTIONunderstand(task) build(context) result = verify() if not result.passed: improve()
04 / EVIDENCE
Test results ✓Protected checks ✓Recorded verdict ✓Known limits ✓
05 / DECISION
EVIDENCE
BEFORE TRUST
Inspect it. Then ship it.

Claims are cheap. Evidence scales.

GitHub ↗

03 / SELECTED WORK

Real problems.
Working systems.

Three builds around one question: how do we make useful AI worth trusting?

View the project index ↓Not just ideas.
Working systems.
— R.V.

Proof of Work

Verification for coding agents.

Proof of Work / in 3 stepsIllustrative demo

agent.patchClaim received

“Done. All checks passed.”

tests/test_result.py− assert result == expected+ assert True

But the agent changed what “passing” means.

An agent submits a patch and claims the task is done.

Problem: An agent’s “done” can hide a broken or weakened check.
My contribution: Built the verification gate and coding-agent evaluation harness.
Stack: Python, SQLite, Ed25519
Evidence: 60 recorded runs across 3 agent configurations on 20 tasks.

Tessera

Local-first AI testing IDE.

Tessera / code → test planIllustrative demo

▧ tessera01 / Your codebaseLocal

⌄ checkout/ cart.ts checkout.ts payment.ts⌄ tests/

checkout.ts

function checkout(cart) {
  validate(cart);
  charge(cart.total);
  return receipt;
}

◉ Source stays on-device

Start with a codebase on your own machine.

Problem: Useful tests need context from the whole codebase.
My contribution: Founded neuratile; built mutation scoring and test-improvement flows.
Stack: Rust, Tauri, React, SQLite
Outcome: Generate QA artifacts; optionally execute and improve tests.

Master Models Beta

Small specialists. Real evaluation.

Master Models BetaIllustrative demo

Small base. Focused roles.01 / Specialize

Qwen34B

FrontendBackendSecurity reviewCode reviewTesting & QA

Five focused coding roles share a small Qwen3-4B base.

Problem: A focused model must earn its place on real tasks.
My contribution: Building frozen evaluations and source-grounded data tooling.
Stack: Python, Qwen3-4B, QLoRA
Evidence: 100 frozen tasks. V2 training and model comparisons pending.

ALSO ON THE BENCH

PR Reliability Platform

Approval-first GitHub review, durable workflows, and commit-bound evidence. Production rollout remains pending.

04 / ARCHITECTURE

Follow the
working parts.

How my systems think, act, and prove their work. Start with the flow; inspect the implementation.

Inputs

Human intent
Source code
Real constraints

Context

Parse + index
Retrieve
Relevant context

Tools

Typed interfaces
Files + search
Execution

Execution

Agent runtime
Bounded steps
Observe

Verification

Tests + checks
Protected gates
Explicit limits

Evidence

Logs + results
Artifacts
Traceability

Release

Human review
Package + ship
Real feedback

↳ Feedback loop · Learn · Improve ↵

View detailed architecture 3 PROJECTS / SOURCE-LINKED MAPS

05 / EVALUATION PLAYGROUND

Small models.
Real work.

100 frozen tasks. Five specialist domains. Real repository changes with checkable criteria.

Master Models · Beta

100FROZEN TASKS

✓ Real repositories
✓ Checkable criteria
✓ Isolated from training

FIVE SPECIALIST DOMAINS

TASK DIFFICULTY

Easy25

Medium45

Hard30

Same tasks.
Inspectable criteria.

Explore the 100 frozen tasks FILTER / INSPECT / EXPORT

Each role has 5 easy, 9 medium and 6 hard tasks.

Frozen source snapshot f16899d ↗. Exports include titles, criteria and source paths; no model outputs or scores.

Evaluation foundation available. Five frontend pilot tasks have execution evidence; independent review remains pending. V2 training and model comparisons have not run. This explorer displays task specifications, not model scores.

06 / INTERACTION LAB

Intent in.
Action out.

Follow the trace, step by step. See the decisions behind the final answer.

Trace everything.
Show the evidence.

CHOOSE A WALKTHROUGH

Simulated steps, real engineering principles. No models or APIs are called.

TRACE / VERIFICATIONReady when you are

// execution trace

›Agent confidence is not an acceptance criterion.
Show me the evidence.

Inspectable steps. Explicit boundaries.Explore the actual project ↗

07 / OPERATING PRINCIPLES

Built a
certain way.

The stack changes. These are the things I keep coming back to.

Context before
cleverness.

Give agents useful retrieval, typed tools, and a clear picture of the problem.

— R.V.

II.

Trust has
a test suite.

✓ Re-run the checks✓ Protect the verifier✓ Inspect the result

“It worked once” is where the work starts.

— R.V.

III.

Own the
whole loop.

Take responsibility from the first brief to the release that reaches someone.

— R.V.

08 / WORKBENCH

Tools for
real work.

A small, purposeful set. Chosen for the work they make possible.

Python

Rust

TypeScript

React

Tauri

SQLite

Docker

Ollama

Tree-sitter

MCP

GitHub Actions

09 / ABOUT

Curiosity is
the constant.

An AI/ML intern at PranavX Labs, based in Solapur, India.

The résumé ↓

Rajveer Vadnal outdoors in Solapur — Rajveer VadnalSOLAPUR, IN

Skeptical of hype.
Including my own.

I’m an AI/ML intern at PranavX Labs and an AI/ML student at SIT Hyderabad. Outside work, I build developer tools at neuratile. I want agents to show their work before anyone trusts the result.

I co-built a municipal WhatsApp assistant. With Proof of Work, I re-run an agent’s checks rather than taking its word for the result.

Diploma completed. Studying AI & ML at SIT Hyderabad through direct second-year admission.

Still learning. Always building.Follow the work ↗

10 / EXPERIENCE + EDUCATION

Work and study.

Where I’ve worked and what I’m learning.

Sep 2026 - Present · 3-month internship

PranavX Labs ↗

AI/ML Engineer (Intern)

Inside the role

Developing PranavX Talent Engine, a local-first resume intelligence and candidate-ranking system using Python, FastAPI, Streamlit, SQLite, and open-source ML models.

Implemented layout-aware PDF/DOCX parsing, skill normalization, and evidence-backed ranking with TF-IDF, sentence-transformer embeddings, FAISS retrieval, and cross-encoder reranking.

Built API-backed recruiter workflows for intake, filtering, review, and export; added model comparisons, CPU benchmarks, and robustness audits with identity-free scoring and source-traceable evidence.

2025 - Present

neuratile (Tessera) ↗

Founder & Lead Engineer

Inside the role

Founded neuratile and built Tessera, a local-first testing IDE using Rust, Tauri, and React. Generation stays on-device with local providers; cloud providers are optional.

Implemented mutation scoring, bounded test-improvement flows, and persisted self-heal history. Tree-sitter and SQLite cosine retrieval supply code context; opt-in Docker runners execute supported tests.

Cross-platform CI checks and release automation target Windows, macOS, and Linux. Installer builds produce drafts; they do not establish published or signed releases.

neuratile on GitHub ↗

2025 - Present

Visage AI ↗

Founder & Lead Engineer

Inside the role

Founded and shipped Visage AI, a mobile product that previews cosmetic procedures on a user’s own face using AI-driven facial analysis.

Took the product from concept to public launch across the mobile experience and live marketing site.

Jun 2025 - Apr 2026

Solapur Municipal Corporation

Software Development Intern

Inside the role

Co-built and shipped a production WhatsApp chatbot giving citizens direct access to government services - currently live in production for the municipal corporation.

Ran stakeholder interviews and cross-department workflow mapping to translate real operational pain points into actionable software requirements.

Delivered production-quality software for a government body at minimal cost as part of a cross-functional engineering team.

2026 - 2029 (Expected)

SIT Hyderabad

Artificial Intelligence & Machine Learning (AIML)

The foundations

Three-year AIML program with expected completion in 2029.

3-year diploma

Government Polytechnic, Solapur

Diploma in Computer Technology

The foundations

Completed in 2026.

11 / CONTACT

Let’s build something real.

Got a hard problem
worth proving?

Agentic AI. Developer tools. A useful idea that needs someone to see it through.

rajveer.r.vadnal@gmail.com ↗

LinkedIn ↗GitHub ↗X ↗Instagram ↗

BUILT WITH INTENTUseful AI.
Human
outcomes.RAJVEER VADNAL / IN

I build systems
that can prove
themselves.

The proof
is in the loop.

Real problems.
Working systems.

Proof of Work

BLOCKED.

Tessera

Master Models Beta

Follow the
working parts.

Small models.
Real work.

Intent in.
Action out.

Built a
certain way.

Context before
cleverness.

Trust has
a test suite.

Own the
whole loop.

Tools for
real work.

Curiosity is
the constant.

Work and study.

PranavX Labs ↗

neuratile (Tessera) ↗

Visage AI ↗

Solapur Municipal Corporation

SIT Hyderabad

Government Polytechnic, Solapur

Got a hard problem
worth proving?

Where to?

Inside the build

Start with the interesting bit.

Notes on this portfolio.

Built to be inspected

The human details

Sources & boundaries

I build systemsthat can provethemselves.

BLOCKED.

Context beforecleverness.

Trust hasa test suite.

Own thewhole loop.

PranavX Labs ↗

neuratile (Tessera) ↗

Visage AI ↗

Solapur Municipal Corporation

SIT Hyderabad

Government Polytechnic, Solapur

Got a hard problemworth proving?

I build systems
that can prove
themselves.

Context before
cleverness.

Trust has
a test suite.

Own the
whole loop.

Got a hard problem
worth proving?