01Distributed LLM Infrastructure2026
InferGate
A distributed LLM inference gateway built for low latency, provider efficiency, and graceful failure.
394ms -> 188ms
p95 latency
sde-1 · systems & applied ai
Gaurav Phadale
Builds systems that have to work.
Distributed infrastructure, agentic AI, and full-stack products. Engineered with evidence behind every claim, not adjectives.
scroll
01 · positioning
Building at the intersection of systems × AI × product.
I care about what happens after the prototype works: latency, reliability, architecture, edge cases, and everything required to turn an experiment into a system somebody can depend on. Most of my work sits where a model, a cache, and a user all have to agree on what “correct” means at the same time.
verified benchmarks
02 · featured work
Systems that survive benchmarks and failure modes.
02Agentic RAG Platform2026
CampusCopilot
A citation-grounded research assistant with hybrid retrieval, calibrated abstention, and reproducible evaluation.
66.4% -> 88.2%
Hit@5
03Industrial Safety Risk Engine2026
Conflux
A real-time safety system that detects compound plant hazards missed by isolated threshold alarms.
10-second
risk-rescore tick
04Walk-Forward Quant Research2026
StockCraft
A no-lookahead equity research pipeline that reports both signal and failure honestly.
54.34% vs 54.29%
directional accuracy vs. baseline
03 · how I think
Systems, not screens.
This is the actual request path from InferGate, one of the four projects below. Hover a stage to see the decision behind it. Every case study gets its own version of this diagram, built from what was actually shipped.
Incoming inference request hits the gateway.
04 · stack
The tools behind the outcomes.
A backend-first stack for building, measuring, shipping, and operating AI products end to end. Hover a skill to see which project actually used it.
Languages & Backend
AI & Retrieval
Data & Infrastructure
Frontend & Product
Engineering Practice
05 · experience
Production work, measured.
From AI product infrastructure to SQL performance and clinical machine learning.
Mentoria
Software Development Engineer I (SDE-1)
Promoted from Full-Stack & Agentic AI Development Intern - Project Arya
- Promoted to SDE-1 after growing AI resume scoring and job-matching to 27,000+ users on arya.mentoria.com during the internship.
- Continuing to own production backend, frontend, and agentic AI workflows across FastAPI, React, PostgreSQL, Redis, and cloud infrastructure.
Mentoria
Full-Stack & Agentic AI Development Intern
Project Arya - AI-powered job search platform
- Grew AI resume scoring and job-matching to 27,000+ users on arya.mentoria.com, shipping semantic search over PostgreSQL + pgvector embeddings and OpenAI APIs in two-week sprints alongside product and design.
- Held p95 latency at 760ms for 50 concurrent users by moving resume parsing, AI inference, and email automation off the request path into Redis-backed Celery/RQ pipelines, with pytest coverage gating every release.
- Built the FastAPI REST APIs behind authentication, application tracking, and Gmail workflows secured with OAuth2 and JWT, plus the React/TypeScript dashboard consuming them.
- Cut recurring hosting spend to zero by migrating 50+ client sites to GitHub Pages and Cloudflare Workers with automated GitHub Actions deploys, and onboarded non-technical clients onto Sanity Studio.
- Shipped a Chrome MV3 extension for job-application autofill and integrated Razorpay payments with role-based access control for enterprise accounts.
SkyHigh Travels
Full-Stack Developer Intern
Production operations and analytics
- Cut average query latency 35% by redesigning the customer, itinerary, and invoice schema and eliminating N+1 access patterns through JOIN restructuring and targeted indexing.
- Shifted operations pricing from intuition-based to data-driven with a SQL analytics module aggregating booking and revenue data by month, segment, and route.
- Delivered production features across a 250-hour engagement through weekly stand-ups with non-technical stakeholders.
NMIMS MPSTME
Student Research Intern - SRIP
Under Dr. Mahesh Patil - 1,166-patient tuberculosis cohort
- Built the full ML pipeline: imputation, normalization, outlier handling, feature engineering, and PCA.
- Beat Logistic Regression and Random Forest baselines with SVM (0.90 AUC, 87% accuracy), isolating alcohol use as the strongest severity predictor across every model.
- Co-authored and presented the resulting paper at the TCSC National Conference, Dec 2024.
Education
NMIMS - Mukesh Patel School of Technology Management & Engineering
B.Tech in Computer Engineering · Aug 2023 - May 2027
Pace Junior Science College, Mumbai
HSC - Science · 2021 - 2023
I.E.S. Secondary School, Mumbai
SSC · 2021
Recognition
AWS Certified Cloud Practitioner
CLF-C02 - Amazon Web Services - 2025
Top 100 of 1,000+ teams
ZS Campus Beats 2026 - ZS Associates
National Conference Presenter
Alcohol Use and Tuberculosis Severity - TCSC - Dec 2024
06 · about
Engineering with evidence, not adjectives.
I’m a Computer Engineering student at NMIMS and an SDE-1 at Mentoria, working across production APIs, retrieval systems, background pipelines, and the React frontends that sit on top of them.
I care about the numbers behind a claim: p95 latency under concurrency, benchmark quality, test coverage, provider-call reduction, or an honest negative result. The goal is never a clever demo. It’s a system whose behavior can be explained and trusted.
Outside of that: usually somewhere between an architecture diagram and a debugger, occasionally a chessboard, and reliably a few chapters into an Agatha Christie novel.
07 · contact
Have a hard problem worth solving?
Tell me what you are building, where it is getting stuck, or what role you have in mind. This form sends directly to my inbox, no email-app redirect needed.
garyphadale@gmail.com