Keshav Kapur outdoors by a canal, smiling

Keshav Kapur

Shipping software that holds up in production.

Software engineer with AI depth, specializing in backend services . Currently at LinkedIn; previously at Shell.

Shipped at

  • LinkedIn
  • Shell
  • Texas A&M
  • ARTPARK

00 · Now

What I’m building this summer.

At LinkedIn I’m shipping production ML services: LoRA fine-tuning, evaluation pipelines, and Flyte workflows that cut inference cost and stay reliable under load. Before that: two years at Shell building .NET services, CI/CD, and Azure infra. Open to full-time software and AI / ML roles, plus Fall 2026 internships.

“Two weeks at LinkedIn down: orientation done, diving into the summer project.” Latest note →

01 · Experience

Where the work ships.

  1. AI / ML Intern

    LinkedIn

    Python · Java · MySQL · PyTorch · Qwen 2.5 · LoRA · Flyte

    • Built production training + evaluation pipelines on Flyte; owned the path from experiment to reliable throughput.
    • Spearheaded LoRA fine-tuning on Qwen 2.5, beating out-of-the-box SOTA models in production eval.
    • Raised model accuracy to 70.4% (vs 49.8% GPT-5.3-Codex, 58% Claude 4.6 Opus).
    • Cut inference cost from $14–$25 to $0.30 per million tokens, a 98% reduction.
  2. Research

    Texas A&M University

    Python · PyTorch · LLMs · DPO · GRPO · SFT

    • Applied DPO, GRPO, and SFT for bias mitigation in a multi-step story generation pipeline.
    • Lifted StereoSet ICAT from 67.5 → 76.97; cut stereotype score toward neutrality.
    • On BOLD, dropped gender bias variance 0.206 → 0.044 and ideology bias 0.515 → 0.183.
  3. Software Engineer

    Shell

    C# · .NET · GitHub Actions · Terraform · Azure · Docker · MySQL

    • Shipped C# / .NET services and APIs that sped internal onboarding by 35%.
    • Automated CI/CD with GitHub Actions & Terraform; provisioned Azure infra for reliable deploys.
    • Built a grounded-retrieval PoC with GPT-3.5-turbo and an Azure-hosted vector index.

02 · Projects

Selected builds from the repo.

Software, systems, and ML work from github.com/Blebot0.

Python · CLIP · FAISS · Fashion Retrieval

Cross-modal search · CLIP + FAISS · query expansion + rerank

Text-to-image search for fashion: CLIP embeddings, FAISS indexing, query expansion, metadata filters, and side-by-side multi-model comparison.

Charts comparing within-community vs cross-community hostility rates on Reddit

Reddit Hyperlink Network

GitHub ↗

Python · Network Analysis · Graph Mining · LIWC · Supervised Learning

8.8% cross-community hostility rate · graph + text features

Data-mining study of the SNAP Reddit hyperlink graph: predicting hostile cross-subreddit links from structure and language.

Recurrent Attention Model

GitHub ↗

Python · PyTorch · REINFORCE · RNN

REINFORCE + supervised hybrid glimpse policy

PyTorch RAM that glimpses image patches like a retina, trains location policy with REINFORCE, and classifies with a hybrid pipeline.

KAN training history curves for loss and metrics over epochs

KAN m-Height Predictor

GitHub ↗

Python · Kolmogorov-Arnold Networks · B-splines · Mixed Precision

B-spline edges · early stopping · full validation suite

Deep KAN regressor that predicts m-height from matrix configurations using learnable edge activations.

CodeWhisper Forge

GitHub ↗

JavaScript · Node.js · Python · Gemini

Success-rate prediction · auto-fix drafts · dry-run PRs

AI issue-resolution assistant that mines similar closed GitHub issues and can generate pull-ready patches.

TAMU Datathon: Product Search

GitHub ↗

Python · BM25 · Embeddings · Cross-Encoder Rerank

Two-stage IR · nDCG / AP / Recall composite

Hybrid BM25 + dense retrieval with cross-encoder reranking for grocery product search.

Behaviour tree diagram for a pick-and-place task with fallback and sequence nodes

LightBT

GitHub ↗

C++ · Behaviour Trees · Robotics · State Machines

Tick-driven sequences · fallbacks · parallels

Lightweight behaviour-tree library for common robot tasks with clear status semantics.

Sparsh

ROS · Python · PyQT · C++ · OCR · Machine Learning

OCR → Braille / voice accessibility hardware

Open-source braille reader: OCR extracts document text, then translates to Braille and voice via physical actuation.

03 · LinkedIn

Recent posts from the feed.

04 · Stack

Languages, infra, and models.

Software

Python · C# · .NET · Java · C++ · JavaScript · FastAPI · React · Docker · MySQL · PostgreSQL · Grafana

Cloud / DevOps

Azure · AWS · Terraform · GitHub Actions · Linux · Kubernetes · Azure DevOps · Flyte

AI / ML

PyTorch · LoRA · RAG · NLP · Deep RL · LLMOps · MCP · AI Foundry

Certifications

Terraform Associate (003) · Azure Fundamentals · Green Software Practitioner

05 · Education

Computer science, end to end.

May 2027

Texas A&M University

Master of Science in Computer Science · GPA 4.0/4.0

Jul 2019 - Jul 2023

Manipal Institute of Technology

BTech Computer Engineering · CGPA 9.1/10 · Rank 7/139 · Head of AI & Automation, Mars Rover Manipal

06 · Publications

Peer-reviewed work.

Comparative Study of Sentiment Analysis on Cyber Security Related Multi-Sourced Data

ATIS 2022 · Springer · BiLSTM F1 0.895

Multi-source Reddit/Twitter sentiment for cyber security attitudes; compared ML, lexicon, and deep models.

DOI ↗

MaNLP@ SMM4H22: BERT for Classification of Twitter Posts

ACL SMM4H 2022 · F1 up to 0.81

Binary age self-report tweet classification for the Social Media Mining for Health shared task.

ACL Anthology ↗

07 · Contact

Let’s build something that ships.

Open to full-time software and AI / ML roles, plus Fall 2026 internships. College Station, Texas. Remote & relocate friendly.