Skip to content
View Ajeenckya5's full-sized avatar
1 follower · 1 following

Block or report Ajeenckya5

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.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. 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
Ajeenckya5/README.md
Typing intro

LinkedIn LeetCode Portfolio Email

🧠 About Me

I'm an AI/ML Engineer with an MS from UW–Madison, specializing in building production-grade LLM systems, agentic pipelines, and RAG architectures from scratch.

  • 🔬 Currently building: self-improving LLM agents with QLoRA fine-tuning and RLAIF feedback loops
  • 🏗️ I write raw API calls over frameworks — my agents don't use LangChain, AutoGen, or CrewAI
  • 🤖 Deep focus on fine-tuning, knowledge distillation, and multi-modal AI
  • 🎯 Open to: ML Engineer · AI Engineer · MLOps · Research Engineer roles

🛠️ Tech Stack

LLMs & Agents

Python PyTorch HuggingFace LlamaIndex Ollama

Vector DBs & RAG

pgvector ChromaDB Mistral 7B LLaMA 3

Computer Vision & Fine-Tuning

OpenCV Vision Transformer QLoRA Fine-Tuning

Backend & Infra

FastAPI Docker PostgreSQL GitHub Actions

🚀 Featured Projects

🤖 Self-Improving LLM Agent

Full self-improving long-horizon agent — strategy memory + distillation

  • Failure-trace analysis → corrective strategies → ChromaDB strategy memory · 90% on Tau Bench
  • Grok-4 teacher labels → QLoRA fine-tunes LLaMA-3.2-1B student
  • 95% inference cost reduction vs teacher at matched quality

QLoRA DPO Knowledge-Distillation ChromaDB LLaMA-3

🖥️ CodeCraft CLI Agent

Production CLI coding agent — zero framework dependencies

  • ReAct loop + 11 workspace tools, raw HTTPS to xAI/OpenAI
  • RLAIF scoring via Grok 4 · JSONL tracing · cross-session memory
  • 7.5× faster than comparable LangChain baseline

ReAct RLAIF Tool-Calling xAI Python

📧 Multi-User Email RAG

E5-small-v2 + Mistral 7B over 10K+ indexed emails

  • JWT auth with SQL-enforced per-user isolation at the pgvector layer
  • Embedding caching + batched retrieval · sub-80ms retrieval latency
  • Precision@5 = 0.84 · Answer faithfulness = 0.79

RAG pgvector Mistral-7B FastAPI JWT

🗺️ Multimodal Optimization Framework

MILP cost–time optimal routing across bus, train & flight networks

  • Exact Mixed-Integer Linear Program: flow conservation, stopover limits, mode feasibility
  • Tested on 2,040 routes from a central U.S. hub
  • Cost–time Pareto trade-offs for multimodal journeys

MILP Optimization Routing Operations-Research NumPy

⚙️ OptiCode AgentR

MINLP-controlled LLM code generation

  • Six-module pipeline selects optimal code candidates
  • Solves a Mixed-Integer Nonlinear Program under latency, cost & accuracy constraints
  • Optimization-meets-LLM: Julia (JuMP) + Python

MINLP Optimization Code-Generation Julia Streamlit

😊 Facial Expression Recognition

CNN vs ViT transfer learning on FER2013 (7-class)

  • 5-model comparative study: CNN → ResNet → ViT → Hybrid → Domain ViT
  • 71.4% accuracy with trpakov/vit-face-expression
  • Full ablation study + per-class precision/recall/F1

ViT CNN PyTorch Transfer-Learning FER2013
➕ More: Operations Research & Optimization

"Build it from scratch. Understand every layer. That's how you ship reliable AI."

🌐 ajeenckya5.github.io — run my models live in your browser

Profile Views

Pinned Loading

  1. codecraft-cli-agent codecraft-cli-agent Public

    Production-grade CLI coding agent built from scratch — ReAct loop, 11 tools, RLAIF scoring via Grok 4, cross-session memory, JSONL tracing. 7.5× faster than LangChain baseline.

    Python

  2. multi-user-email-rag multi-user-email-rag Public

    Multi-user email RAG: E5-small-v2 + Mistral 7B GGUF over 10K+ emails, JWT auth, SQL-enforced per-user isolation (pgvector cosine), FastAPI backend.

    Python

  3. facial-expression-recognition facial-expression-recognition Public

    FER2013 comparative study: CNN baseline vs ViT transfer learning vs CNN-Transformer hybrid. ViT (trpakov/vit-face-expression) outperforms from-scratch CNN on 7-class emotion recognition.

    Jupyter Notebook

  4. multimodal-optimization-framework multimodal-optimization-framework Public

    Constrained multimodal routing: MILP-based cost-time optimal paths across bus, train & flight networks. Tested on 2,040 routes from a central U.S. hub.

    Python

  5. self-improving-llm-agent self-improving-llm-agent Public

    Self-improving long-horizon LLM agent — ChromaDB strategy memory + failure analysis, Grok-4 teacher labels → QLoRA-distilled LLaMA-3.2-1B student. 90% on Tau Bench, 95% inference cost reduction.

    Python

  6. opticode-agentr opticode-agentr Public

    MINLP-controlled LLM code generation. Six-module pipeline selects optimal candidates by solving a Mixed-Integer Nonlinear Program under latency, cost & accuracy constraints.

    Python