Hero Background Composition Madhav Kannan K Portrait
Open to internships

Hi, I'm Madhav Kannan K

MSc Data Science student
building ML systems

Building computer vision for infrastructure inspection, ML diagnostic prototypes, document search, and open-source tooling.

Scroll down
Background & Competencies

About & Skills

I am an MSc Data Science student at Coimbatore Institute of Technology (CIT). My coursework and projects focus on machine learning modeling, computer vision, data pipelines, and developer tooling.

I focus on building practical systems: training defect segmentation models on infrastructure benchmarks, building semantic vector search over heterogeneous documents, and contributing upstream patches to open-source developer toolchains.

Education
MSc Data Science · Coimbatore Institute of Technology · 2024–2029
Certification
Mathematical Foundations of Machine Learning (IISc Bangalore, NPTEL) · View Certificate
Languages
English (Professional Working Proficiency), Malayalam (Native), Tamil (Native)

Programming

Python / Go (Golang) / SQL

Machine Learning & CV

PyTorch / scikit-learn / XGBoost / OpenCV / SHAP

Backend & Tooling

FastAPI / Docker / Linux / Bash / Git & GitHub

Frontend & UI

React / JavaScript / HTML5 & CSS3

Upstream & Community Code

Open Source

#322
google-ml-infra / rules_ml_toolchain MERGED

Windows Python Discovery & Sequential Fallback Probing

Fixed hermetic Bazel build initialization failures on Windows hosts where system Python is invoked as python.exe or via the py launcher rather than python3.

Bazel / Starlark / Python Toolchain / Windows Runners

View Merged PR #322
#769
ollama / ollama-python OPEN

Two-Way Tool Correlation & ToolCall ID Preservation

Fixes dropped tool call results between Ollama's Go server backend and Python client by retaining correlation IDs during Pydantic serialization.

Python / Pydantic / Ollama SDK / Tool Calling

View Open PR #769
fafa
Madhavkannank / fafa OWN PROJECT

Go Port of GoogleChromeLabs/jsbi (ECMAScript BigInt)

Go stdlib math/big.Int lacks ECMAScript BigInt semantics: it misses NaN relational guards and does not invert negative shift counts (x << -y == x >> y). See the repository for other documented edge cases and the full 30-bit limb engine design.

9,696,250 differential fuzz cases (88.7% statement coverage) · Benchmarks: Multiply 94.88 ns (3-limb operands), Divide 304.9 ns (4-limb by 2-limb operands)

Go (Golang) / BigInt Arithmetic / Differential Fuzzing / V8 Equivalence

View Repository & Benchmarks
Systems & Models

Featured Projects

COMPUTER VISION · PYTORCH

KCI — Kandhari Construction Intelligence

AI infrastructure inspection prototype for bridge structural defect detection and severity assessment.

42.15% mIoU
4 Classes
975 Val Images
14.2 ms Forward Pass (only)
Defect Class Validation IoU
Crack 34.58%
Rust 48.86%
Spalling 46.48%
Exposed Rebars 38.69%

4-class subset; also used for checkpoint selection; not comparable to the 19-class dacl10k leaderboard.

Problem: Manual civil infrastructure inspections are slow, subjective, and lack reproducible defect segmentation.
Approach: Trained an FPN with an EfficientNet-B4 backbone on 640×640 bridge imagery and paired it with a rule-based AASHTO condition state evaluation engine.

Python · PyTorch · segmentation-models-pytorch · OpenCV · FastAPI · React

+ Architecture Pipeline Deep Dive
01

FPN + EfficientNet-B4 Inference (models/fpn_efficientnet.py)

Processes 640×640 civil bridge inspection imagery through a Feature Pyramid Network. Segments 4 primary defect classes with cross-entropy and Dice loss optimization.

02

AASHTO Condition State Scoring (backend/risk_engine.py)

Evaluates segmented mask areas against AASHTO condition state standards (CS1–CS4) using rule-based threshold formulas without cloud dependencies.

03

Automated PDF Audit Generator (backend/report.py)

FastAPI endpoint renders defect overlays, priority scores, and condition breakdowns into downloadable engineering PDF summaries.

BIOINFORMATICS · TRANSCRIPTOMICS In development

GUIDO — Genomic Unified Intelligence for Disease Oncology

Transcriptomics cancer biomarker modeling and literature linking prototype connecting ranked genomic candidates to PubMed abstracts (Groq LLaMA 3).

Problem: High-dimensional RNA-seq expression data suffers from extreme feature collinearity and high false-discovery rates.
Approach: Training regularized XGBoost models on TCGA cohorts from UCSC Xena, evaluating fold-level SHAP feature stability, and linking ranked biomarkers to PubMed abstracts using Groq LLaMA 3.
Status: Active development on reproducible preprocessing, cross-validation and TreeSHAP stability pipelines; research prototype, not clinically validated.

Python · XGBoost · SHAP · scikit-learn · Biopython · FastAPI · Docker

+ Architecture Pipeline Deep Dive
01

TCGA RNA-Seq Ingestion (data/cohort_loader.py)

Standardizes normalized gene expression data from UCSC Xena across TCGA cancer cohorts, applying variance thresholds and log2 transforms.

02

Regularized XGBoost & SHAP Stability (models/risk_model.py)

Computes TreeSHAP attribution scores across validation folds to isolate stable biomarker candidates and suppress fold-specific noise.

03

PubMed Abstract Retrieval (audit/ncbi_entrez.py)

Queries NCBI Entrez for published literature on ranked genes and feeds retrieved abstracts to Groq LLaMA 3 for citation-backed synthesis.

APPLIED NLP · VECTOR SEARCH

AXEL — Automated Document Extraction & Search

Document ingestion, OCR, and semantic vector embedding pipeline for multi-format text retrieval.

Problem: Extracting and searching information across heterogeneous files (PDFs, scans, DOCX) is slow when limited to exact string matching.
Approach: Combined Tesseract OCR and pdfplumber for text extraction, generated 384-dimensional dense embeddings with sentence-transformers, and injected metadata tags back into file properties.
Result: Enables natural language semantic search and metadata tag injection across multi-format documents (PDF, DOCX, and images).

Python · sentence-transformers · Tesseract OCR · pdfplumber · ChromaDB

+ Architecture Pipeline Deep Dive
01

Multi-Format Parsing & OCR (pipeline/ingest.py)

Parses text from PDFs and DOCX files; applies Tesseract OCR to image files and scanned pages.

02

Dense Vector Embedding (models/embeddings.py)

Chunks extracted text and computes normalized 384-dimensional embeddings (all-MiniLM-L6-v2) for cosine similarity ranking.

03

File Attribute Injection (pipeline/tagger.py)

Injects extracted keywords into EXIF headers and document properties, keeping search tags portable with the file.

SYSTEMS & AUTOMATION

Jarvis Jr — Multimodal Desktop Assistant

Local Python desktop assistant runtime with validated tool execution, voice routing, and recovery journaling.

360+ Automated Tests
Problem: General desktop assistant wrappers lack path validation boundaries, local tool verification, and offline math calculation.
Approach: Built an intent router and tool dispatcher with canonical filesystem path boundary validation, SymPy symbolic math solving, and wake-word voice handling.
Result: Runs locally offline across Fedora Linux and Windows with local GPU and VRAM monitoring; validated across 360+ automated test cases covering semantic intent routing, path boundary validation, and transactional recovery journaling (tests/).

Python · Pydantic · SymPy · openWakeWord · OpenCV · PyAutoGUI

+ Architecture Pipeline Deep Dive
01

Semantic Intent Classification (core/router.py)

Routes incoming spoken and text inputs into application commands, calculations, or local queries without sending voice data to cloud providers.

02

Security & Boundary Validation (security/permission_engine.py)

Enforces canonical path resolution, symlink boundary checks, and permission gating prior to file read/write or application launch commands.

03

Action Journaling & Retry (recovery/engine.py)

Maintains transactional execution logs, rolling back transient state changes and retrying failed commands up to configurable limits.

More Projects

Audit-QP

FastAPI application evaluating examination papers against syllabus modules and Bloom's taxonomy distributions using spaCy text parsing and semantic embeddings.

Python · FastAPI · spaCy · sentence-transformers · PyMuPDF Code
AURA-ECE

Transforms classroom teacher observations into developmental insights using voice/text input (Whisper via Groq), Presidio PII anonymization, MongoDB storage, and a Streamlit dashboard.

Python · FastAPI · Streamlit · MongoDB · Groq Code
Color Suggestion App

Java Swing desktop application for creating, visualizing, and organizing color palettes with relational persistence.

Java · Swing · MySQL · JDBC Code
BeatDrop

Distributed Big Data analytics platform processing 239,000 tracks across a 5-tier HDFS lakehouse with MapReduce, Pig ETL, Hive data warehousing, and PySpark.

Hadoop HDFS · MapReduce · Apache Pig · Apache Hive · PySpark · MongoDB
Leadership & Operations

Experience

2026 – Present
FOSS CIT · Open Source Student Chapter

Joint Secretary

Coordinate core operations, event logistics, and technical initiatives for the student open-source chapter.

  • Organized a 3-day Linux workshop and framed the technical workshop syllabus for participating students.
  • Drive student engagement in regional FOSS initiatives, mentoring peers on version control standards, code reviews, and open-source licensing.
2025 – 2026
FOSS CIT

Technical Team Member & Event Lead

Organized technical event logistics, problem tracks, and automated evaluation pipelines for competitive student developer events.

  • Event Lead — Nakshathra 2026: Directed challenge logistics and flow across 4 simultaneous technical competitions on a single day for second-semester engineering students, introducing them to practical coding challenges.
FEATURED EVENT LEAD

I'M GOING INN (IGI) Hackathon

Led technical operations for a 2-day story-driven web application competition with 100+ participants, deploying an automated Gemini-assisted grading pipeline for submission evaluation.