SHAYAR SHRESTHA
0%

// Software Engineer · AI Engineer · Builder

SHAYAR
SHRESTHA

M.S. Computer Science · 4.0 GPA · Germantown, MD

_

· OPEN TO WORK · AVAILABLE ·

SCROLL

LangChain
FastAPI
Go
.NET Core
Kafka
PostgreSQL
LLM Evaluation
RAG Pipelines
OpenTelemetry
Docker
Redis
React
TypeScript
Python
Claude API
OpenAI
LangChain
FastAPI
Go
.NET Core
Kafka
PostgreSQL
LLM Evaluation
RAG Pipelines
OpenTelemetry
Docker
Redis
React
TypeScript
Python
Claude API
OpenAI
01

// 01. About

I build things
that run.

I'm a software engineer and AI systems builder with 4+ years of production experience and an M.S. in CS (4.0 GPA, Hood College), graduated January 2026. Most of my career has been on systems that actually had to keep working — high-volume pipelines, production APIs, multi-agent platforms.

I don't deploy things and hope for the best. I instrument them, measure them, and fix what's actually broken. Recently I've gone deep on AI engineering: LangChain agents with tool-calling and safety guardrails, RAG pipelines, LLM evaluation frameworks, and the observability you need to trust that an agentic system behaves correctly at 2am when you're not watching.

Open to SWE, AI Engineer, ML Engineer, Data Scientist, and Research roles.

0
YRS PRODUCTION EXP
4.0
GPA · M.S. CS
0
RECORDS / RUN
0
LLM PROVIDERS
Available for new roles
LOCATION
Germantown, MD
DEGREE
M.S. CS · Hood College
GRAD DATE
January 2026
CERT
Azure AZ-900
PRIMARY
Python · Go · C#
OPEN TO
Onsite (Relocate) / Hybrid / Remote
INTERESTS
Agentic AI Systems LLM Evaluation Medical AI RAG Pipelines Distributed Systems Data Engineering
02

// 02. Experience

Where I've
shipped things.

01
Graduate Assistant — IT Applications
Hood College · Frederick, MD
Jan 2025 – Dec 2025
+
  • Departments like the Registrar, Financial Aid, and Grad School were stuck in a loop of manual SSRS exports and IT-tweaked stored procedures for every data request. I identified the real problem and built the solution: an ASP.NET MVC platform with scheduled jobs and role-based access that eliminated the dependency entirely.
  • Pulled execution plans for ~35 stored procedures, found root causes systematically — bad joins, table scans, misplaced business logic — and fixed them. Performance improved ~6%, with tuning ongoing.
  • Worked directly with non-technical department heads throughout. That tight feedback loop between technical and non-technical people is often where the real requirement lives.
ASP.NET MVCSQL ServerTypeScriptSSRS~6% perf gain
02
Software Engineer
UBA Solutions — Offshore for Monotype Imaging, USA
Apr 2020 – Nov 2023
+
  • Built a .NET data ingestion pipeline (RabbitMQ, S3, React UI) for onboarding product catalogs from newly acquired companies. Each source arrived in different formats with different data quality problems. Ingestion went from ~3 months down to ~1 day for up to 10,000 products per run.
  • Browse and search traffic was hitting the database directly and slowing the platform during peak usage. Designed a dedicated service with Redis caching, materialized views, and TTL rules by data type. Database load dropped ~30%, API response time improved ~40%.
  • Drove an incremental React migration of a legacy JavaScript storefront, measuring Core Web Vitals at each step with Lighthouse. Steady, measurable progress without a risky full rewrite.
.NET CoreRabbitMQRedisReact3mo → 1 day-30% DB load+40% API perf
03
Associate Software Engineer
Sursa Technology · Early-Stage Startup
Aug 2018 – Oct 2019
+
  • Built .NET Core REST APIs for a user management platform — OAuth 2.0, JWT, Google Sign-In, role-based access control, middleware-level auth enforcement. Security fundamentals established here that I've built on ever since.
  • Containerized the service with Docker. Regular code reviews and design discussions with senior engineers. Where I learned to care about getting the details right.
.NET CoreOAuth 2.0JWTDockerREST APIs
03

// 03. Projects

Things I built because
they needed to exist.

01

// FOUNDING ENGINEER · 2025 – PRESENT

LinguaFlow AI

Production multi-agent platform I designed and built as sole engineer. Real-time Kafka pipeline → LangChain agents with tool-calling, per-session memory, multi-step reasoning, and safety guardrails. Agents would sometimes produce responses that looked correct but were wrong in subtle ways — I built evaluation logic that catches those before anything reaches a user. Full observability: OpenTelemetry, Prometheus, Grafana. Measurement is how you know whether a system is actually working.

Next.jsFastAPIGoLangChainKafkaPostgreSQLRedisDocker
linguaflow.ai
AGENTS: 12
LATENCY: 84ms
ERRORS: 0.2%
SESSIONS: 847

02

// FOUNDING ENGINEER · 2025 – PRESENT

Tell Me More

I built this because most LLM deployments lack the evaluation infrastructure to detect subtle reliability problems before they affect users. Routes identical prompts across OpenAI, Gemini, and Anthropic. Logs quality, latency, and token cost per run. The core is the evaluation schema — structured A/B comparisons across model versions and prompt templates, designed to distinguish whether a change actually improved output or just shifted the failure mode.

FastAPILangChainPostgreSQLPydanticReact
deepoctopus.com
GPT-4
94ms
Claude
71ms
Gemini
118ms
// 2,847 evaluations logged

03

// BACKEND ENGINEER + PRODUCT OWNER · 2024 – 2025

Certification Helper App

Full-stack certification exam prep platform. C# / .NET Core backend with CQRS and Domain-Driven Design, Entity Framework Core, PostgreSQL, React + TypeScript on the frontend. Also led the team as Product Owner — managing the roadmap and architecture alongside doing the engineering. Where I got the most hands-on CQRS and DDD practice in a real project rather than just an exercise.

C#.NET CoreCQRSDDDEF CorePostgreSQLReact
DOMAIN
APP
INFRA
API
04

// 04. Stack

The stack,
honestly.

Production tools I use daily. Honest about what I'm strong in vs. familiar with. The orbital visualization on the right shows the full picture — inner ring is where I live, outer is where I'm growing.

AI & Agents
LangChainRAGOpenAIClaude APIGeminiPrompt Eng.LLM EvalMCP ToolsGuardrails
Backend & Systems
Python / FastAPIGoC# / .NET CoreTypeScriptREST APIsGraphQLCQRSOAuth 2.0
Data & Infra
PostgreSQLMongoDBRedisKafkaDockerAzureOpenTelemetryGrafana
ML & Research
TensorFlowKerasHugging FaceScikit-learnPyTorchNLPFine-tuning
STACKSHAYAR
05

// 05. Writing

What I've been
thinking about.

Medium Mar 2, 2026 4 min read

What I Learned From Reading the DistilBERT Paper

BERT is powerful but slow and expensive to run in production. The DistilBERT paper introduced knowledge distillation — a smaller student model learns not just from labels, but from how the full teacher model behaves. Three combined loss signals. 40% fewer parameters. 60% faster. 97% of BERT's performance. This paper shifted how I think about practical NLP systems: it's not just about removing parameters — it's about transferring understanding.

BERTDistilBERTNLPAILLM
Read on Medium →
Code Maze Published

C# Local Functions: An In-depth Guide

Local functions look like a minor syntactic convenience until you use them properly. This article covers when to reach for them over lambdas, how they interact with closures and captured variables, and the performance implications that actually matter in production. Published on Code Maze, one of the largest .NET engineering publications.

C#.NETBackend
Read on Code Maze →
✍️

MORE ON MEDIUM

AI systems, NLP research, and the practical side of building things that work in production — not just in notebooks.

Follow on Medium →
06

// 06. Contact

Let's build
something real.

I'm open to SWE, AI Engineer, ML Engineer, Data Scientist, and Research roles. If you're working on something interesting and need someone who ships things that actually work in production — let's talk.

[email protected]