I build AI systems and agentic applications with real engineering constraints: tools, memory, checkpoints, evaluation, inference, latency, and cost — not just chat wrappers.
My work spans LLM agents, AI infrastructure, inference optimization, evaluation, and developer tooling, with a strong focus on understanding why a system works, where it fails, and how to make it more reliable and efficient.
Resume · Portfolio · LinkedIn · Microsoft Learn · LeetCode · Codeforces · Kaggle · Email
This year: 186 contributions · 20 commits · 13 PRs · 1 review
Achievements: Pull Shark · YOLO · Quickdraw
I contribute to AI infrastructure, agent runtimes, inference systems, and developer tooling across several open-source projects.
- Microsoft Agent Framework — agent workflows, persistence, checkpoints, and runtime behavior
- ONNX Runtime — inference/runtime work including quantization-related paths
- LangGraph — Pregel / DeltaChannel persistence, checkpoint reconstruction, HITL workflows
- AutoGen — multi-agent workflows involving planning, tools, and memory
- Future AGI — evaluation and AI tooling
Upstream repositories:
ONNX Runtime · LangGraph · AutoGen · Microsoft Agent Framework · Future AGI
| Project | What I built / explored |
|---|---|
| Apertus Eval Prep | Reproducible LLM evaluation across models, prompts, inference configurations, and hardware |
| InferLite | Hardware-aware inference optimization using latency, throughput, memory, and Pareto analysis |
| IndicQuant | Efficient evaluation and optimization for Indic-language document AI |
| Agent Systems | Tool calling, memory, checkpoints, RAG, HITL, and workflow orchestration |
| Code Execution Platform | Multi-language code execution and backend infrastructure |
I work across the stack when the problem requires it:
| Area | Tools |
|---|---|
| LLMs / ML | PyTorch, Transformers, Hugging Face, LoRA, QLoRA, quantization |
| Agents | LangGraph, AutoGen, Microsoft Agent Framework, tool calling, checkpoints, RAG + FAISS |
| Inference | ONNX Runtime, vLLM, llama.cpp, GPTQ, AWQ |
| Backend | FastAPI, Flask, PostgreSQL, Redis, Node.js |
| Systems | C++, Python, Linux, Docker, CI/CD |
| Data / Search | PostgreSQL, MongoDB, FAISS, ChromaDB |
| Cloud | AWS, GCP, Vercel |
I prefer technologies based on the engineering problem they solve, rather than collecting tools for a stack list.
For example:
- FastAPI → lightweight async APIs for model and agent serving
- PostgreSQL → durable, queryable application and agent state
- Redis → low-latency ephemeral state and caching
- FAISS → efficient local vector retrieval
- ONNX Runtime → optimized and portable model inference
- Docker → reproducible deployment environments
- C++ → performance- and systems-sensitive components
- LangGraph / Agent Framework → explicit stateful workflows rather than opaque agent loops
For flagship projects, I document the complete engineering lifecycle:
Problem → Existing limitations → Architecture → Engineering decisions → Agent workflow → Tools → Failure modes → Evaluation → Performance → Cost → Deployment → Future improvements
The goal is not simply to make an AI system work.
It is to understand:
Why does it work? Where does it fail? What trade-offs did I make? And how do I know the system is actually better?
I enjoy algorithmic problem solving and systems-oriented programming.
ICPC Online Challenge 2026 — Rank 1864
LeetCode — 1708 · 525+ problems solved
Codeforces — Shivani_103
- Efficient LLM inference
- LLM evaluation and reliability
- Quantization and model optimization
- AI agents and agent runtimes
- AI for Software Engineering
- Multilingual / Indic LLMs
- AI infrastructure
- Research engineering
Microsoft Learn · Responsible Generative AI · GSSoC