Vaibhav Mahore

AI/ML EngineerBangalore, India

VaibhavMahore

Building reliable AI systems, efficient sequence models, and evaluation infrastructure for intelligent software.

View ProjectsView Resume
GitHubLinkedIn

01/Introduction

I work across the full arc of an AI system: architectures from first principles, evaluation that proves they behave, agents that finish with green tests.

I'm Vaibhav, an AI/ML engineer and B.Tech student in Mathematics and Computing at the Indian Institute of Science. Recent work spans benchmarking state-space models against Transformers at Ericsson, reproducing Mamba-3's capability claims in pure PyTorch, and building a reliability engine for LLM/RAG systems.

The common thread: measure first, then believe.

efficient sequence modeling (SSMs)LLM evaluation & observabilitytool-using agents & sandboxingreproducible experimentation

Location
Bangalore, India
Education
IISc · B.Tech Mathematics and Computing
Focus
Sequence models · LLM evaluation · reliable agents
Recently
AI/ML Research Intern, Ericsson

02/Interactive lab

Watch the systems run.

Play, step, and reset each one - teaching schematics on top, measured numbers underneath.

01/ 03sequence modeling · architectures

Efficient Sequence Modeling: From S4 to Mamba-3

Five sequence architectures implemented from first principles, trained at matched budgets, and measured against their own papers' claims - OOM points included.

0.983

parity accuracy, Mamba-3 (baselines 0.51-0.56)

Explore full case studyGitHub
x1x2x3x4x5x6x7x8x9x10x11x12hₜ = Ā·hₜ₋₁ + B̄·xₜ → yₜ = C·hₜchunk 1→chunk 2→chunk 3→chunk 4chunked_scan()block matmuls · T/64sequential_scan()reference loop≡atol 1e-314-test suite · aligned + misaligned chunksS4DLTI · convMambaselectiveMamba-2chunked SSDMamba-3complex Δ

Tokens inA sequence of tokens enters the model. Every architecture in the study reads this same input through one shared skeleton.

01 / 08

03/Experience

Three roles, one practice: make the claim measurable.

01/Research

AI/ML Research Intern

Ericsson India Private Limited · Bangalore

May 2026 - Jul 2026

  • Benchmarked WiMamba vs. Transformer on AWS A10G GPUs, proving O(T) linear scaling with 10.7× lower latency (30ms vs 327ms) and 42.6× less GPU memory (113MB vs 4.8GB) at 4×4 patch size.

  • Evaluated geographic OOD transfer (Florida → San Diego) across 5 tasks, achieving a 0.9626 composite score (+16.6% over baseline) and 7.4–7.7 dB lower NMSE on channel estimation and interpolation.

  • Explored a Soft Unfreezing Parameter-Efficient Fine-Tuning strategy using PyTorch linear probing, freezing 75% of parameters while retaining 98.5% of full fine-tuning performance.

WiMambaOOD TransferParameter-Efficient Fine-TuningPyTorch

02/Industry

Machine Learning Engineer (AI Data Trainer)

Alignerr · Remote

May 2025 - Aug 2025

  • Authored structured reasoning traces and high-quality training data to improve LLM decision-making, tool selection, and step-by-step problem-solving across complex technical tasks.

LLM Training DataReasoning Traces

03/Industry

Machine Learning Engineer

micro1 · Remote

May 2025 - Aug 2025

  • Contributed high-quality data annotation, evaluation, and model validation workflows for frontier AI systems.

  • Performed data preprocessing and supported machine learning model development using TensorFlow and Scikit-learn.

Data AnnotationModel ValidationTensorFlowScikit-learn

04/Selected work

Selected systems I built, tested, and verified.

Three deep dives. Every claim traces to a benchmark, a test suite, or a public repository.

01/ 03sequence modeling · architectures

Efficient Sequence Modeling: From S4 to Mamba-3

Minimal S4D, Mamba, Mamba-2 and Mamba-3 implemented in pure PyTorch from first principles, with the benchmarks to show exactly where selective state spaces win.

  • Implemented minimal S4D, Mamba, Mamba-2 and Mamba-3 from first principles, including a chunked State-Space-Duality path test-verified against a sequential scan.

  • Benchmarked latency, memory and throughput at 1K–32K sequence lengths on a 6 GB RTX 4050.

  • Reproduced the Mamba-3 paper's capability claim: complex-valued transitions solve parity tracking (98.3%) where real-transition SSMs and Transformers stay at chance (≤56.3%); on selective copying, LTI S4D collapses (6%) vs 67% for Mamba-1/2.

98.3%
parity tracking solved
≤56.3%
real-transition / transformer baseline
1K–32K
sequence lengths benchmarked
6 GB
RTX 4050, full test suite

Python · PyTorch · NumPy · SciPy · Hugging Face · Matplotlib · pytest

Explore case studyLive demoGitHubTechnical report
S4DLTI · CONVOLUTIONMambaSELECTIVE SSMMamba-2SSD · CHUNKEDMamba-3COMPLEX-VALUED1K32K CONTEXTBENCHMARKED AT EVERY OCTAVE · RTX 4050 · 6 GBSELECTIVE COPYING · TEST L=128LTI S4D 6%SELECTIVE SSM 67%

02/ 03LLM evaluation · reliability

AIRE: AI Reliability & Evaluation Engine

A trace-based evaluation engine that turns LLM/RAG behavior into replayable evidence: deterministic metrics, a failure taxonomy, and direction-aware regression detection.

  • Built replayable run storage, deterministic metrics (groundedness, citation coverage, retrieval recall/MRR), a failure taxonomy, and a direction-aware regression engine with per-case attribution.

  • Evaluated three RAG versions over a 26-case suite, flagging an intentional fluency-for-grounding regression (groundedness 0.57 → 0.00).

  • A live Gemini responder then fixed abstention (0.0 → 1.0) and grounding (1.0) at 15 s/case.

26
evaluation cases
3
RAG versions compared
0.57→0.00
regression caught
15 s
per case, live responder

Python · pytest · HTML/CSS

Explore case studyLive demoGitHubTechnical report
INPUTRETRIEVALEVALUATIONREGRESSIONATTRIBUTIONv2 · 0.57v3 · 0.00GROUNDEDNESS · FLAGGED REGRESSIONABSTENTION 0.0 → 1.026-CASE SUITE · 15 S/CASE · LIVE RESPONDER

03/ 03autonomous agents · verification

Autonomous Repository Engineer

A verified tool-using coding agent: the planner only proposes actions; a deterministic state machine decides, and completion requires green test evidence, never model claims.

  • A deterministic state machine (PLAN → ACT → OBSERVE → VERIFY) wraps the planner, which only proposes actions.

  • Seven validated tools with schema checks, a workspace path jail, AUTO/ASK/DENY permission policies and command allowlists.

  • Benchmark: 4/4 SWE tasks deterministically with zero failed tool calls; a live Gemini planner passes 2/2 on the same runtime.

4/4
SWE tasks, deterministic
0
failed tool calls
7
validated tools
2/2
live Gemini planner

Python · pytest · Git

Explore case studyLive demoGitHubTechnical report
SANDBOX BOUNDARYPLANACTOBSERVEVERIFYSTATEMACHINE7 VALIDATED TOOLS

05/Research

Questions I keep returning to.

Work-in-progress notes rather than publications. Every entry below links to a public repository on this page.

  1. R.01

    S4 → Mamba-3, implemented from first principles

    Minimal S4D, Mamba, Mamba-2 and Mamba-3 in pure PyTorch with a test-verified chunked SSD path: complex-valued transitions solve parity tracking at 98.3% where LTI S4D collapses to 6%.

    Independent · s4-to-mamba

  2. R.02

    Evaluation as replayable evidence

    A trace-based engine for LLM/RAG reliability: deterministic metrics, a failure taxonomy, and direction-aware regression detection with per-case attribution.

    Independent · AIRE

06/Capabilities

A toolkit, organized by what it builds.

Languages
PythonC++Basic SQL
AI / ML
PyTorchTensorFlowScikit-learnNumPySciPyPandasHugging Face DatasetsMatplotlibpytest
LLM & AI Systems
LLM evaluation & observability (tracing, judges, regression testing)RAGtool-using agentssequence modeling (Transformers, State Space Models)
MLOps & Cloud
DockerAWS (EC2, S3)LinuxGit/GitHub
Core Concepts
Deep LearningModel ArchitectureExperiment Design & ReproducibilitySoftware Engineering Practice (testing, sandboxing, CI-style verification)

07/Foundations

The base the work stands on.

Education

Indian Institute of Science, Bangalore

B.Tech in Mathematics and Computing · Aug 2023 - Present

Relevant coursework

  • Data Structures and Algorithms
  • Linear Algebra
  • Probability
  • Machine Learning
  • Deep Learning
  • NLP

Prior: St. Ann's Senior Secondary School, CBSE · Class XII 2023 · Class X 2021

Certifications

Oracle Professional

August 2025

Leadership

Sci-Tech Coordinator · Rhapsody Fest, IISc

2023 - 2026

  • Led end-to-end organization of three inter-college gaming tournaments and a competitive coding contest for 200+ registered teams.
  • Administered a 500+ member community for registration and tech support, and drove fest publicity through digital campaigns.

08/Next

Let's build something intelligent.

mvaibhav@iisc.ac.in
GitHubLinkedInResume+91-9284782375