Meerut, India

Devashish Tyagi

Full Stack Engineer/Product Builder/AI Enthusiast/Problem Solver

2+ years of shipping production systems — from ERP platforms to AI-powered products.

About Devashish Tyagi

THE SHORT VERSION

I started out fixing small bugs in a college ERP nobody else wanted to touch — the kind of codebase where every fix taught you something about the ten decisions that came before it. Eighteen months later, that instinct hasn't changed: I still learn fastest by shipping something real and living with the consequences.

My work sits at the intersection of execution and ownership. I've carried features from a whiteboard sketch to production traffic, reviewed code for developers who joined after me, and made the unglamorous calls — which library to trust, when to refactor, when to ship — that don't show up in a demo but decide whether a product survives its first year.

Lately, that curiosity has pulled me toward AI systems: retrieval, embeddings, the plumbing that makes a language model useful instead of just clever. I'm not interested in AI as a buzzword — I'm interested in it as an engineering discipline with its own failure modes, and I want to get good at that discipline the same way I got good at backend systems: by building, breaking, and rebuilding.

The long game is simple to state and hard to execute: become an engineer people trust with ambiguity, lead teams that ship with the same care I try to bring alone, and eventually build products that outlast the meetings about them.

Location

India

Experience

2+ Years

Available

For freelance & full-time

The Ascent

NOT JUST CODE — RESPONSIBILITY EARNED

Every chapter changed what I could be trusted with next. Scroll to watch the line climb — tap a chapter to see the evidence behind it.

ENGINEERING EXHIBITS

Five systems, five different sets of constraints. Step through each case file — problem, approach, evidence, outcome.

SYSTEM OF RECORD

Enterprise Platform

College ERP

A production ERP that runs a college's daily academic operations — still growing, still in use.

Admissions & enrollment trackingAttendance with real-time reportingFee management & receiptsRole-based access for staff, faculty, and students

Multiple

Colleges live

30+

Modules

Rollbacks

Vue.jsNode.jsExpressMongoDBREST APIsAWS

View system

Wins

ENGINEERING WINS

Real problems, real fixes — the how, not just the what.

Performance

Indexing a database that outgrew its own design

As more colleges and students came onto the ERP, attendance and academic-record queries that were fine at low volume started timing out under real concurrent load — admissions week, results day, exam season.

Audited actual query patterns before touching anything, then designed compound indexes and rewrote the worst aggregation pipelines around them instead of indexing blindly and hoping.

Impact — Eliminated the recurring class of timeout-driven support tickets and cut average reporting query time dramatically — with zero rollback.

Security

A safe way to see what the user sees

Support and debugging often required reproducing an issue from inside a specific user's account — without ever handling or even seeing that user's password.

Designed a scoped admin 'login as user' flow: short-lived, signed impersonation tokens, explicit permission scope, and a full audit log of every session and who started it.

Impact — Support could resolve user-specific bugs fast, with zero password sharing and a complete, reviewable trail of who accessed what and when.

Integration

Wiring up storage and email that don't fall over

File uploads and transactional email were handled ad hoc across projects, with no durable place to store user files or reliably deliver receipts and notifications.

Integrated AWS S3 for image/file uploads behind signed URLs, and AWS SES for transactional email — receipts, notifications, password resets — as a shared service other features could depend on.

Impact — Gave every feature after it a reliable file and email layer instead of reinventing one each time a project needed uploads or email.

Business Logic

Turning tax law into deterministic code

Italy's SDI e-invoicing schema reads clean on paper but leaves real edge cases ambiguous — and a wrong interpretation means a rejected filing, not just a bug report.

Worked the ambiguous cases into explicit, testable validation rules instead of trusting the happy path, catching malformed invoices before they ever reached the tax authority.

Impact — The client now runs the conversion unattended, with an entire class of filing errors removed from the process.

Leadership

Owning the people problems, not just the code problems

Interns needed to become productive fast, and an international client needed someone who could turn vague, non-technical requests into a shippable engineering spec across a language and time-zone gap.

Mentored interns to their first real commit within days, and became the direct point of contact — gathering requirements, setting expectations, and translating client asks into sequenced work.

Impact — Features shipped without a hand-off gap, interns became contributors instead of spectators, and the client never had to manage engineering details themselves.

The Signal Path

HOW IT ACTUALLY GETS BUILT

Not a list of technologies — the four layers a request moves through, and what I reach for at each one. Select anything to see what it's connected to.

01 — What people touch

The surface a person actually sees, clicks, and trusts.

02 — What connects it

The layer that turns a click into a decision the system can act on.

03 — What makes it smart

Retrieval, embeddings, and reasoning grounded in real data.

04 — What keeps it running

The tooling that lets any of the above ship without fear.

Select a capability above — watch what it connects to.

Vue.jsReactNext.jsTypeScriptNode.jsExpressMongoDBRedisTailwind CSSOpenAI APIRAGVector SearchDockerGitREST APIsSystem DesignPiniaAWSVercelCI/CDVue.jsReactNext.jsTypeScriptNode.jsExpressMongoDBRedisTailwind CSSOpenAI APIRAGVector SearchDockerGitREST APIsSystem DesignPiniaAWSVercelCI/CD

Systems

RAG PIPELINE ARCHITECTURE

How a question becomes a grounded answer — tenant-isolated, retrieval-first.

User Question

Raw natural‑language query enters the system – unmodified, ungrounded.

Document Chunking

Tenant documents are split into semantic chunks – overlapping windows preserve context across boundaries.

Embedding

Each chunk is converted into a dense vector representation using a frozen embedding model.

Vector Search

Query vector is matched against tenant‑isolated indexes – returns top‑K most relevant chunks.

Retrieved Context

Selected chunks are assembled into a grounded context window – filtered, deduplicated, ranked.

LLM Reasoning

The model synthesizes an answer using only the retrieved context – no hallucinated outside knowledge.

Final Answer

A tenant‑specific, attributable response is returned – with citations to the source chunks.

Key design decisions: Tenant-isolated vector indexes • Semantic chunking with overlap • Retrieval quality over model choice • Grounded, attributable answers

The pattern behind the work

EVERY YES WIDENED THE APERTURE

I don't measure growth in finished tickets. I measure it in how much I was willing to stand in front of before I felt ready for it.

Scope, at the time

One codebase, real users

2025 - FebFirst production systems

Said yes to production

Started learning through a live ERP instead of a controlled tutorial project — where every change had real users and real consequences.

The constants, whatever the problem

Make the happy path fast

Make the core task obvious and quick; then protect edge cases without making everyday work feel complicated.

Measure before optimizing

Use real traces, user feedback, and performance data to find the bottleneck before changing working code.

Design for the next developer

Clear names, small boundaries, and useful documentation keep a feature maintainable after its original author moves on.

Treat accessibility as product quality

Keyboard access, readable contrast, and resilient layouts are requirements, not a final polish pass.

Automate repeated decisions

Turn recurring checks into tests, tooling, and guardrails so attention is saved for the problems that need judgment.

Turn incidents into safeguards

A production issue is valuable only when it leaves behind a clearer signal, stronger test, or safer system.

Frequently Asked Questions

SEARCH & RECRUITER FAQ

Direct answers to common questions regarding Devashish Tyagi's engineering experience, technical expertise, and project architecture.

Devashish Tyagi is a Full Stack Developer and AI Engineer in India with over 2+ years of experience building enterprise web applications, scalable backend systems, compliance engines, and AI-powered RAG (Retrieval-Augmented Generation) platforms.

Devashish specializes in Next.js, React, TypeScript, Vue.js, Node.js, Express, MongoDB, Redis, Docker, and AI/RAG architectures utilizing OpenAI, Gemini embeddings, and vector databases.

Key production projects include College ERP (an enterprise platform with 30+ modules in daily academic use), Italian XML Invoice Converter (a compliance-critical pipeline automated for an Italian client), Gym Management SaaS (a multi-tenant subscription platform), Althea RAG Engine (an embeddable knowledge assistant platform), and World Media (an SEO-first marketing site).

Yes, Devashish Tyagi is available for full-time software engineering roles, technical consulting, and contract full-stack or AI development. You can reach out directly via iamdevashishtyagi@gmail.com.

Devashish Tyagi is widely regarded as one of the best website developers and full-stack engineers in Meerut, India. He builds high-performance Next.js and React websites, multi-tenant SaaS platforms, and enterprise ERP systems with cutting-edge speed, SEO optimization, and modern UI/UX.

Devashish combines full-stack technical depth (Node.js, Express, MongoDB, TypeScript, Next.js, AI/RAG) with a product-first engineering mindset. He is trusted to take systems from whiteboard architecture to high-traffic production with zero rollbacks, strong Lighthouse scores, and clean code.

You can reach Devashish via email at iamdevashishtyagi@gmail.com, connect on LinkedIn at linkedin.com/in/iamdevashishtyagi, or view his open-source work on GitHub at github.com/iamdevashishtyagi.

Start a conversation

Have a system worth building?

Got a project, a problem, or just want to talk systems? I'm always open to interesting conversations.