Open to new opportunities

Nikhil
Agrawal

Applied AI Engineer · RAG, Agents & AI Evaluation

I build production retrieval and agent systems, and measure — with statistics, not vibes — whether they actually got better. That's an evidence-grounded RAG and agent platform orchestrated with LangGraph and governed through MCP, a canonical data platform spanning 1.7M+ records, and local LLM & GPU systems work below the API layer.

View Projects Resume Contact
GitHub LinkedIn AIthors
~1,600
civic meetings indexed & evaluated
retrieval and evaluation in production
6
production systems shipped at OpGov
owned end-to-end, from ingestion to evaluation
1.7M+
profiles unified
canonical public-record platform
3rd
MCP × A2A Hackathon
prize winner
Selected Projects

Featured Work

Production-grade AI systems built for scale, reliability, and real-world impact.

Evidence-Grounded Civic Intelligence Platform (ASK NOHA)
Hybrid RAG, knowledge-graph and agent retrieval over ~1,600 civic meetings — orchestrated with LangGraph, exposed as governed tools through MCP, with human approval required on sensitive actions.
Python
LangGraph
MCP
FastAPI
MongoDB
Neo4j
Weaviate
FAISS
Gemini
CrossEncoder

→ Agentic retrieval with governed tools, human approval, and enforced abstention

AI Evaluation & Regression Harness
A versioned evaluation harness with real statistics behind it — bootstrap confidence intervals on retrieval deltas, ablation studies, and an LLM-as-judge calibrated against human labels via Cohen's kappa.
Python
Bootstrapping
Ablation Studies
LLM-as-Judge
Cohen's Kappa
Langfuse
OpenTelemetry

→ Statistical rigor as a release gate: bootstrap CIs, ablations, calibrated LLM-judge

Multi-County Public Records Data Platform
A canonical MongoDB platform unifying 1.7M+ public-record profiles and ~4M historical participation records from incompatible county schemas.
Python
MongoDB
PyMongo
ETL
Identity Resolution
Temporal Data

→ 1.7M+ profiles · 20,060 new profiles reconciled, 1 ambiguous identity isolated

Geospatial Field Operations Dashboard
A live field dashboard over a structured voter database — GPS tracking, nearest-neighbor search over 2dsphere-indexed GeoJSON, and optimized multi-stop routing. Used in a real campaign.
Next.js
TypeScript
MongoDB
GeoJSON
Google Routes API
Python

→ Live dashboard used in a real campaign, with optimized multi-stop routing

MCP × A2A Hackathon — Multimodal Agent System
A multimodal autonomous agent system for complex task automation, built against the Model Context Protocol and agent-to-agent communication. 3rd Prize.
MCP
A2A
Multimodal
Autonomous Agents

→ 3rd Prize — MCP × A2A Hackathon

GPU & Local Inference Lab
A systems-focused lab of six experiments on local LLM inference, GPU memory, scheduling and fine-tuning — measuring what most demos skip.
PyTorch
CUDA
Local LLMs
Quantization
QLoRA
GPU Profiling
Benchmarking

→ Six experiments: inference, memory, scheduling & fine-tuning

Pet POV Wearable — Edge AI Device
A founder's early concept turned into a working prototype: a wearable capturing point-of-view video with on-device YOLO object detection and multimodal emotion inference.
Raspberry Pi
Pi AI Camera
YOLO
TensorFlow Lite
Python
React

→ 0→1 hardware + edge AI product, from founder concept to working prototype

View All Projects
Philosophy

How I Think About AI

Principles that guide every system I build.

Systems, not demos

I focus on AI systems that survive real-world messiness: noisy inputs, partial failures, edge cases, and changing data formats.

Feedback loops matter

The best AI systems are not one-shot generators. They generate, evaluate, improve, and become more reliable over time.

AI is infrastructure

The next frontier is better systems around models: observability, reliability, cost control, and trust at scale.

Currently interested in

Retrieval quality
Structured outputs
AI evaluation
Observability
Local LLM inference
GPU memory & quantization
Scalable backend design
Full-stack AI products
"I build AI systems that don't just work once — they keep getting better."
Learn More About Me