Skip to content
View Keerthirajan58's full-sized avatar
💭
Pursuing my Passion
💭
Pursuing my Passion
4 followers · 11 following

Achievements

Achievement: Pull SharkAchievement: Quickdraw

Achievements

Achievement: Pull SharkAchievement: Quickdraw

Highlights

  • Pro

Block or report Keerthirajan58

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
keerthirajan58/README.md
A terminal window. Left, a dot-matrix portrait of Keerthirajan. Right, a WHOAMI list with the handle @Keerthirajan58: Keerthirajan Senthilkumar; AI Engineer at EventDone; previously ML Engineer Intern, Apple Siri; based in San Francisco, CA, open to relocate; M.S. Computer Science, GWU 2026; focus on LLM evals, agents and LLM infra; shipping bugbout, open source on PyPI; stack: Python, TypeScript, Go, SQL, PyTorch, Transformers, LoRA/PEFT, LiteLLM, Ollama, RAG, MCP, FastAPI, Docker and Kubernetes; links: keerthi58.vercel.app, keerthirajans@gwu.edu and LinkedIn /in/keerthirajans58. The status bar reads: Open to AI/ML and SWE roles, San Francisco, Pacific time. A terminal prompt types one line and stops: Building bugbout: which coding agent really fixes bugs?

Hi there, I'm Keerthirajan 👋

🤖 AI Engineer at EventDone: I deployed a local LLM alongside paid cloud models behind one LiteLLM gateway, with multi-model routing and failover
🍎 Previously an AIML Machine Learning Engineer Intern at Apple (Siri Core Modeling), where I built a 1,096-query benchmark for ambiguous tool calls
🐛 Building bugbout, an open-source CLI on PyPI that finds out which coding agent actually fixes the bugs in your codebase
🏆 2nd place at the Dell x NVIDIA AI Hackathon (Night Shift) · Best Hack for Social Good at Technica 2025 (BarterBrAIn)
🎓 M.S. in Computer Science, The George Washington University (2026)
🚀 Open to AI/ML and software engineering roles · San Francisco, CA, open to relocate
📫 Reach me at keerthirajans@gwu.edu or check out my portfolio

LinkedIn   Portfolio   Resume (PDF)   Email   bugbout on PyPI   Medium

Featured work 🛠️

bugbout. Open source on PyPI. Replays a repo’s own fixed bugs to find out which coding agent actually fixes them. 27 of 66 “done” patches were wrong; hidden fail-to-pass tests; MIT-licensed Python CLI. Night Shift. 2nd place at the Dell x NVIDIA AI Hackathon. An on-call agent built in one day, with all LLM inference local on an NVIDIA GB10. 4 parallel LLM specialists; fixes need approval in Slack; 384 tests. When to act, when to ask. Research with Llama 3.1 8B and LoRA. Fine-tuned an LLM to call the right tool or ask for clarification when a request is ambiguous. Accuracy 60.9% to 70.9%; clarification 80.5% to 95.1%; 110-query held-out set. Voice-agent evaluation harness. Evaluation with the OpenAI Realtime API. A speech-to-speech caller that placed 100 scripted calls to an LLM voice agent. 14 verified bugs; Twilio Media Streams and FastAPI; 482 tests at 98.7% coverage. Ambiguous tool calling. Apple Siri Core Modeling internship. A 1,096-query benchmark across 8 API domains that tests when an LLM should act or ask. Gorilla OpenFunctions v2 accuracy fell from 83.5% to 33.3% when queries were ambiguous; LLM-as-judge labelling; 731 failures in 5 types. BarterBrAIn. Best Hack for Social Good at Technica 2025. A campus barter app with Gemini-powered valuation and negotiation. Gemini 2.5 multimodal APIs; a confidence score per valuation; Flutter and Firebase.

Open source 🌱

My open-source activity from GitHub and PyPI, refreshed weekly: contributions in the last year, pull requests merged upstream, pull requests in review, upstream repositories and the latest bugbout release, then my upstream pull requests, newest first, each with its status (such as Merged or In review), repository, summary and month. A text version is linked below.

Pull requests: nebari-dev/llm-serving-pack#179 · openteams-ai/collab-hub-pack#79 · nebari-dev/nebi#567 · ShishirPatil/gorilla#1365 · nebari-dev/data-science-pack#246 · nebari-dev/llm-serving-pack#186

Text version of this card

Signals 📡

Two charts. Work by area: a radar chart of how many of my 13 roles and projects since 2025 involve each area (some cover several): Full-stack and APIs 6, Evaluation 5, Agents and tool use 5, LLM infra 3, ML and vision 3, Fine-tuning 2. Languages: a bar and a list of each language's average share of code per public repository, the top five languages and Other, with the date of the weekly update and the number of repositories. A text version is linked below.

Text version of these charts

What do I know? 📜

Skills: Python, TypeScript, JavaScript, Go, C++, Bash, React, Next.js, Tailwind CSS, PyTorch, FastAPI, Django, Flask, Docker, Kubernetes, PostgreSQL, Azure and AWS

Also: SQL · Hugging Face Transformers · LoRA/PEFT · SFT · LangChain · LiteLLM · Ollama · RAG · MCP · OpenAI Realtime API · LLM-as-judge · pytest

Pinned Loading

  1. bugbout bugbout Public

    Find out which coding agent actually fixes bugs in your codebase.

    Python 3

  2. nightshift nightshift Public

    On-call AI agent that detects service failures, investigates them and proposes fixes an operator approves in Slack. All LLM inference runs locally on an NVIDIA GB10. 2nd place, Dell x NVIDIA AI Hac…

    Python

  3. Tool-Ambiguity-in-LLMs Tool-Ambiguity-in-LLMs Public

    When to act, when to ask: a benchmark for ambiguous tool calls, and Llama 3.1 8B fine-tuned with LoRA to ask for clarification instead of guessing.

    Python

  4. Keerthirajan-AIML-Fluid-Portfolio Keerthirajan-AIML-Fluid-Portfolio Public

    Personal portfolio of Keerthirajan Senthilkumar, AI/ML and software engineer. Next.js, TypeScript, Tailwind CSS.

    TypeScript

  5. nebari-dev/llm-serving-pack nebari-dev/llm-serving-pack Public

    Self-hosted LLM serving via llm-d. Includes per-model OIDC access control, API key management, and Envoy AI Gateway token counting and rate limiting.

    Go 5 3

  6. Pretty-Good-AI-Ai-Engineering-Challenge Pretty-Good-AI-Ai-Engineering-Challenge Public

    Voice-agent evaluation harness: a speech-to-speech caller on the OpenAI Realtime API placed 100 scripted calls to an LLM voice agent and surfaced 14 verified bugs.

    Python