Shivam Thakur

Professional Experience

Production AI deployments & enterprise manufacturing systems

Held end-to-end technical ownership across the company's core AI, software, and digital manufacturing initiatives. Led the architecture, engineering, and factory-floor rollout of five production systems—spanning edge acoustic quality assurance, private generative AI tooling, high-volume warehouse logistics traceability, and real-time cloud IoT telemetry.

5 Systems Led Architecture & Plant Rollout

100.0% Defect Recall on Factory Line

100k+ / day WMS Traceability Records

~45ms Shop-Floor Edge Inference Latency

Flagship Deployment · Physical AI & Edge DSP • Guided by CTO Tejas Jayaraman & Dr. Veena Mayya (MIT Manipal)

✓ Operational on Factory Floor

AI-Based Acoustic Motor Fault QA & XAI Pipeline

Automated acoustic quality assurance system for automotive electric window motors. Converted 8-second operational sound cycles into 224×224 Mel-spectrograms via STFT, benchmarked 6 architectures to resolve a 1D embedding bottleneck, and deployed a MobileNetV2 classifier with Grad-CAM explainability achieving 100% defect recall at ~45ms edge latency.

Fault Recall 100.0% 38/38 Defective Motors Caught (0 Escapes)

Edge Latency ~45ms MobileNetV2 on Factory Floor Compute

Validation Acc 84.62% Outperformed CRNN (78.5%) & XGBoost (68%)

Industrial Dataset 312 Units Semi-Anechoic Chamber @ Plant

Production Apparatus & Explainability Artifacts (Click to Enlarge)

Testing Environment — Semi-anechoic enclosure with GRAS 46AE mic & pneumatic jig at manufacturing plant.

Grad-CAM Explainability — Proving model activates on true bearing impulses rather than factory ambient noise.

Architecture Benchmark — Visual-acoustic 2D spectrograms resolved the 1D latent embedding bottleneck.

Physical Data Acquisition: Designed the shop-floor testing protocol in a semi-anechoic enclosure using a calibrated GRAS 46AE free-field microphone, capturing 8-second forward and reverse cycles under mechanical load.
Visual-Acoustic DSP Pipeline: Converted 44.1 kHz WAV recordings into 224×224 log-Mel spectrograms via STFT (2048 window, 512 hop, Hann window, 128 Mel bands) to isolate abnormal bearing rotation (ABR) and internal wear conditions (WC).
Edge Architecture Benchmarking: Resolved a 1D embedding bottleneck (where VGGish/YAMNet plateaued at 56–72% accuracy) by fine-tuning MobileNetV2 with spatial transfer learning, achieving 84.62% multi-class accuracy, 0.84 F1, and 100% defect recall at ~45ms inference latency.
Gradient-Based Explainability (XAI): Integrated Grad-CAM heatmaps to verify that neural activations localized to genuine bearing raceway impulse frequencies rather than transient factory noise.

PyTorch MobileNetV2 Audio DSP STFT Log-Mel Grad-CAM XAI Librosa FastAPI Optuna

Proprietary Industrial IP

Internal Enterprise LLM Platform & CAR Quality Portal

Deployed a private, air-gapped LLM platform using Open WebUI and local Ollama inference serving Llama 3, Gemma, and Qwen models with custom company knowledge stores.
Built a full-stack Corrective Action Request (CAR) web portal to digitize quality defect workflows and root cause analysis (RCA) tracking across manufacturing teams.

Open WebUI Ollama Llama 3 FastAPI Docker

High-Throughput WMS Traceability & BOM Engine

Engineered barcode scanning ingestion and dynamic linking logic reconciling two separate warehouse management systems: Aditya Auto and German Tier-1 partner Edscha.
Processed 100,000+ daily transaction records, resolving multi-level child-to-parent assembly lineage across incomplete legacy Bills of Materials (BOM) with zero line stoppages.

Python SQL ETL Pipelines Data Reconciliation Docker

Full-Stack Azure IoT Cloud Telemetry Platform

Built an end-to-end FastAPI + React real-time air purifier telemetry platform deployed on Azure Virtual Machines for plant-floor device monitoring.
Prototyped Azure IoT Hub and Device Provisioning Service (DPS) architectures before adapting ingestion pipelines for constrained industrial network bandwidth.

FastAPI React Azure IoT Hub Azure DPS REST APIs

Automotive OEE Analytics & Automated Production ETL

Architected production PowerBI dashboards with advanced DAX measures tracking Overall Equipment Effectiveness (OEE) and line utilization for German supplier Edscha.
Built automated Python ETL pipelines with custom pairing logic to clean, impute, and reconcile missing legacy sales and operational manufacturing records.

PowerBI Advanced DAX Python Data Modeling ETL

Projects & Research

Selected AI architectures, peer-reviewed research, and open-source tools

Multimodal Audio AI · VIGAM Labs • Sep 2026 – Nov 2026

9.5ms CLAP ONNX + HNSW ✓ 1.0 Beta Live

KLANG — Cross-Modal Audio Retrieval & Synth Morphing Engine

Offline-first desktop audio search and synthesizer preset engine built at VIGAM Labs with Tauri (Rust), React, and a quantized Dual-Tower CLAP INT8 ONNX + USearch HNSW backend. Searches 15,800+ local audio samples and Vital, Surge XT, and u-he patches by natural language, voice, or acoustic similarity in ~9.5ms on CPU.

Interface & Architecture Previews (Click to Enlarge)

Cross-Modal Search & Macro DSP Filters — Querying 'warm analog pad' in ~9.5ms with real-time waveform peaks, musical key badges, and perceptual tone/attack dials.

Synth Preset Inspector & Audio-to-Vital Morphing — Inspecting ADSR envelopes, filter cutoff, drive saturation, and native OS Drag-to-DAW export.

KLANG Retrieval Latency Percentiles Benchmark

Production Latency Benchmark (15,818 HNSW Vectors) — 9.52ms p50 text-to-audio search, 7.19ms audio similarity, and 0.359ms USearch ANN traversal.

Search Latency (p50) 9.52 ms 0.359ms HNSW Lookup (15.8k Vectors)

Retrieval Accuracy 100% R@5 0.958 MRR Across 6 Sound Domains

Peak Throughput 157.2 QPS 322 MB Total Disk Footprint (CPU)

Key Detection Gain +126% N=8192 FFT & EDMA Profiles (~9ms)

Engineered a quantized Dual-Tower CLAP (ONNX Runtime INT8) and memory-mapped USearch HNSW retrieval engine indexing 13,588 audio clips and 2,230 synth presets into a shared 512-D latent space with 9.52ms median latency and 157.2 QPS throughput.
Resolved RoBERTa BPE subword modifier hijacking via 3-prompt canonical ensembling, composite SQLite covering-index hybrid recall, and head-noun lexical re-ranking—lifting multi-modifier Precision@10 from 10% to 100%.
Re-architected musical key detection with N=8192 high-resolution FFTs (2.69 Hz sub-bass bin spacing), sub-semitone tuning-cent compensation, and Sha'ath EDMA profiles, boosting exact tonic accuracy by +126% (19% → 43%, 63%+ harmonic compatibility) in ~9ms pure NumPy.
Built a native Tauri v2 desktop shell (Rust parallel Rayon/Walkdir crawler + OS-level Drag-to-DAW integration for Ableton/FL Studio) alongside an acoustic-to-parameter regression pipeline morphing raw audio into 100% schema-valid Vital, Surge XT, and u-he Diva patches.

Python ONNX Runtime (INT8) CLAP USearch HNSW Audio DSP Rust / Tauri v2 React & TypeScript FastAPI SQLite

Live Product Site ↗

VIGAM Labs ↗

Agentic AI / Local RAG • Apr 2026 – Present

100% Local Inference ● In Active Development

SILO — Local Agentic RAG Platform

On-premises, privacy-first Retrieval-Augmented Generation platform built with FastAPI, LangGraph, Qdrant, and Ollama. Executes stateful Corrective RAG (CRAG) and Self-RAG grading loops with hybrid vector retrieval and real-time SSE citation streaming.

Interface & Architecture Previews (Click to Enlarge)

Collections & Knowledge Management — Uploading files, configuring chunking, and tracking Qdrant vectorization.

Multi-Chat Session History — Real-time token streaming with inline cited sources and optional web search fallback.

Engineered a 100% local Agentic RAG backend integrating FastAPI, Qdrant hybrid search (dense + BM25 sparse vectors), and Ollama (Llama 3.1/3.2) with zero external data transmission.
Architected cyclic state machines in LangGraph executing Corrective RAG (CRAG) and Self-RAG loops with automated document relevance grading and query rewriting to suppress hallucinations.
Integrated SentenceTransformers cross-encoder re-ranking, DuckDuckGo web search fallback, and Server-Sent Events (SSE) token streaming with inline source citation mapping.
Containerized multi-service orchestration via Docker Compose with persistent volumes for document stores and Qdrant vector indices.

Python FastAPI LangGraph Qdrant Ollama PyTorch Docker

Source Code on GitHub ↗

Computer Vision / IEEE Access Publication • Sep 2024 – Feb 2025

IEEE Access (Vol. 13, 2025) ✓ Published

VIGAM — Drone Super-Resolution & Disease Diagnosis

Coupled Real-ESRGAN and Vision Transformer (ViT) pipeline restoring degraded low-altitude drone imagery across 54,634 samples (38 classes) for automated plant pathology diagnosis. Peer-reviewed and published in IEEE Access.

Classification Acc 90.0% ViT 38 Classes (54,634 Images)

Gain vs Baseline +15.4% Outperformed Swin (+5.9% vs ViT)

Perceptual Quality 0.1501 LPIPS 21.97 BRISQUE Restoration Score

Co-authored peer-reviewed paper in IEEE Access (Vol. 13, pp. 66432–66444, 2025): 'Leveraging Super-Resolution Technology in Drone Imagery for Advanced Plant Disease Diagnosis and Prognosis'.
Engineered a coupled Real-ESRGAN (515 layers, 16.7M parameters, 57 Residual-in-Residual Dense Blocks) and Vision Transformer pipeline to recover high-frequency lesion textures from degraded aerial imagery.
Achieved 90.0% multi-class classification accuracy across 38 categories, outperforming standalone ViT baselines by 5.9% and Swin Transformer variants by 15.4%.
Integrated a fine-tuned GPT-2 module to generate automated pathology reports and actionable treatment prescriptions from classification outputs.

PyTorch Real-ESRGAN Vision Transformers (ViT) GPT-2 Albumentations OpenCV

Read Paper on IEEE Xplore (DOI) ↗

Mobile Systems / Sensor & Audio DSP • Aug 2026 – Present

Kotlin & Jetpack Compose ● In Active Development

Caliber — Mobile Hardware Diagnostics & Sensor Suite

Open-source Android hardware diagnostics and sensor calibration suite built in Kotlin and Jetpack Compose. Bypasses VSYNC frame-batching to profile sub-frame touch digitizer polling (240Hz–720Hz+), 400Hz+ 6-axis IMU drift, and 48 kHz acoustic loopback latency.

Interface & Architecture Previews (Click to Enlarge)

Touchscreen & Digitizer Lab — Real-time standard deviation jitter tracking and uncoalesced sub-frame event capture.

Controller Precision Lab — 360° circularity error benchmarking, deadzone mapping, and Hall Effect vs. potentiometer verification.

Dual Theme Architecture — Pure OLED pitch-black gaming mode vs. daylight review interface with state-preserving bento dashboard.

Touch Digitizer 240–720Hz+ Uncoalesced MotionEvent Profiling

IMU Telemetry 400Hz+ DSP Zero-Point Drift & 3-Axis Oscilloscope

Acoustic Loopback 48 kHz Chirp Cross-Correlation Audio Latency

Built an uncoalesced touch-digitizer profiler reading raw historical MotionEvent buffers ahead of VSYNC frame delivery, isolating true hardware polling rates (240Hz–720Hz+) and spatial jitter.
Implemented a high-frequency SensorManager pipeline (SENSOR_DELAY_FASTEST, >400Hz) with a custom Compose DrawScope 3-channel rolling oscilloscope and one-tap zero-point tare calibration.
Designed a gamepad stick circularity benchmark computing 360-sector radial error and deadzone hysteresis to distinguish Hall Effect sensors from worn potentiometers.
Engineered a sample-accurate acoustic loopback engine synthesizing Hann-windowed 2.4 kHz audio chirps at 48,000 Hz with time-domain cross-correlation delay detection.

Kotlin Jetpack Compose Android SensorManager AudioRecord DSP MotionEvent APIs DrawScope Canvas

Source Code on GitHub ↗

Technical Skills & Education

Core AI frameworks, backend engineering & academic credentials

Languages

Python, C++, Rust, TypeScript, JavaScript, Kotlin, SQL, HTML/CSS

AI & Machine Learning

PyTorch, ONNX Runtime (INT8), TensorFlow, Scikit-learn, XGBoost, Computer Vision (OpenCV, ViT, GANs), Audio DSP, XAI (Grad-CAM)

GenAI, Multimodal & Vector Search

LangGraph, Agentic RAG (CRAG, Self-RAG), CLAP, USearch HNSW, Qdrant, Ollama (Llama 3, Qwen, Gemma), SentenceTransformers

Backend, Desktop & Cloud

FastAPI, Flask, React, Tauri v2, SQLite, REST APIs, Docker, Linux, Git, Azure (VMs, IoT Hub), GCP (Vertex AI), ETL, Power BI (DAX)

Education & Certifications

MIT Manipal (MAHE) 2024 – 2026

M.Tech in Artificial Intelligence & Decision Science

CGPA: 8.29 / 10.0 · Manipal, India

SASTRA Deemed University 2020 – 2024

B.Tech in Computer Science & Engineering

First Class with Distinction · Thanjavur, India

Google Cloud Skills Boost Verified

Machine Learning & Generative AI Badges

Vertex AI Prompt Design Transformers & BERT Responsible AI

Beyond Tech

Creative pursuits, indie game prototyping & community

Writing & Music Production

When I'm away from models and code, I dive into creative writing and music production—penning long-form game critiques and industry essays, and producing lofi and retrowave tracks inspired by video game soundtracks.

Indie Game Prototyping

Experimenting with gameplay mechanics, state machines, and spatial audio in Godot and Unreal Engine. Currently co-developing a multiplayer co-op horror platformer with a friend in my spare time.

Gaming & Community

Avid fan of narrative RPGs, tactical stealth, and indie titles. Always happy to connect and chat about games, soundtracks, or design!

Steam: MavihsLH ↗ Discord: mavis_lh

Professional Experience

Aditya Auto Products and Engineering

AI-Based Acoustic Motor Fault QA & XAI Pipeline

Internal Enterprise LLM Platform & CAR Quality Portal

High-Throughput WMS Traceability & BOM Engine

Full-Stack Azure IoT Cloud Telemetry Platform

Automotive OEE Analytics & Automated Production ETL

Projects & Research

KLANG — Cross-Modal Audio Retrieval & Synth Morphing Engine

SILO — Local Agentic RAG Platform

VIGAM — Drone Super-Resolution & Disease Diagnosis

Caliber — Mobile Hardware Diagnostics & Sensor Suite

Technical Skills & Education

Beyond Tech

Writing & Music Production

Indie Game Prototyping

Gaming & Community

Let's build something impactful.