▸ Founder, Tulips.edu · AI & Backend Engineer · CS + AI @ Plaksha

Swapneel.

builds systems for when things break — not just when they work.

  • engineer by day,
  • philosopher by night,
  • musician by choice,
  • rebel by fate.
See the work

3 schools in production·sole engineer @ flyback·300+ students onboarded to AWS

scroll to wander

while I breathe, I hope —

dum spiro, spero.

I'm Swapneel — a CS + AI student at Plaksha University who ended up running production software before finishing college. I founded Tulips.edu, a school ERP that three schools use every day, and I'm the one who gets the 2 a.m. page when something breaks.

That changed how I build. Demos run on the happy path; the real world doesn't. So I design for the failure first — offline queues, fallbacks, guardrails, confidence scores instead of confident guesses.

Away from the terminal I wander: mountains when I can reach them, books when I can't. I write poems and short stories, debate philosophy, look up at the stars more than is reasonable, am learning guitar, and I've been a Poké-nerd since the DS days.

now —

  • founder Tulips.edu · since Jun '26 · I own uptime, backups, onboarding and the pager
  • flyback Founder's Office · since Aug '26 · sole engineer on the platform and staff CRM

recently —

  • aws Campus Ambassador · Aug–Sep '26 · onboarded 300+ students to AWS Builder
  • research RA with Prof. Anuj Kapoor (Univ. of Missouri) · May–Aug '26 · causal inference

roles — and the pagers that came with them

experience.

EXP—01

in production

Tulips.edu

Founder · Jun '26 — now · Ranchi · school ERP, live in 3 schools and used daily

Founded and built it alone, and I own the whole stack: architecture, deployment, operations, plus client onboarding, pricing, support and incident response. Multi-tenant school ERP: student records, fees (UPI deep links + QR), academics and parent communication across three surfaces — public site, staff ERP, parent portal — on one SPA. Tenant isolation is enforced in middleware with tenant_id on every table; every API route is role-checked across five roles; every state change emits an immutable audit event.

failure-mode thinking

The most frequent failure in an Indian school isn't the server — it's the teacher's connectivity. Attendance is offline-first: writes queue in IndexedDB and background-sync on recovery. Uploads go direct to R2 via presigned URLs, so storage failures never touch the app server. Migrations are raw versioned SQL, because schemas holding fee records don't get ORM surprises.

  • FastAPI · Python 3.11
  • PostgreSQL + asyncpg
  • Preact + TypeScript
  • Docker Compose
  • nginx
  • Cloudflare + R2
visit tulipsedu.in ↗
Tulips.edu — production topology, as deployed ALL SYSTEMS OPERATIONAL
teachers' phones
cloudflare
indexeddb queue

Knock out api, postgres or r2. The log says what actually happens in production — including the one failure with no fallback.

EXP—02

sole engineer · now

Flyback

Founder's Office Intern · Aug '26 — now · recruiting platform matching overseas Indian professionals with high-growth Indian teams

The only engineer on the team, leading architecture and infrastructure. Rebuilt the production platform and the staff CRM from scratch, and built the sourcing side: custom scrapers and ATS ingestion pipelines that keep the candidate pool fed automatically.

failure-mode thinking

Matching runs in two passes because either one alone fails: a deterministic arithmetic role matcher is precise but brittle, semantic ranking is flexible but drifts. Human-in-the-loop validation queues sit on top, so the taxonomy doesn't quietly rot.

  • Full-stack rebuild
  • Hybrid matching
  • Semantic ranking
  • Human-in-the-loop
  • Scrapers · ATS ingestion
visit flyback.in ↗

EXP—03

300+ onboarded

Amazon Web Services

Campus Ambassador · Aug — Sep '26 · Plaksha University, Mohali

Contracted to drive student adoption of AWS developer and cloud-learning programmes on campus. Ran tabling activations, technical walk-throughs and digital outreach, and beat the regional campaign quota by onboarding 300+ students to AWS Builder platforms.

  • Developer outreach
  • Technical walk-throughs
  • AWS Builder
  • Campus campaigns

EXP—04

causal inference

University of Missouri

Research Assistant · May — Aug '26 · with Prof. Anuj Kapoor · Continuity of Care on Digital Health Platforms

A large-scale digital physiotherapy dataset. When a physiotherapist exits unexpectedly, it breaks a patient's care mid-episode. These provider-side shocks are plausibly exogenous — clean variation for asking how much continuity causally matters, and for whom.

My work spanned the identification strategy and the estimation pipeline, including pressure-testing parallel-trends assumptions across two difference-in-differences designs and validating dose-response gradients.

the hard part

The treatment is a path through an adaptive system, where standard regression breaks down.

identification
Staggered DiD (Callaway–Sant'Anna), provider exits as instruments, discrete-time hazard models.
causal ml
Double/debiased ML, causal forests, R-/DR-learners for heterogeneous effects across patient fragility.
sequence models
Transformer encoders over care trajectories, counterfactual RNNs, representation balancing (TARNet/CFRNet).

built to survive contact with production

projects.

PRJ—01

open source · RAG

CitefyMe

local-first research RAG where every claim carries its evidence — and a verdict

Ask questions over papers pulled live from arXiv (abstracts, or whole papers via HTML/PDF ingestion). The pipeline retrieves with BM25, Gemini embeddings or hybrid RRF, extracts each claim from the answer, and runs it back against its cited passage for a supported / partial / unsupported verdict. Hand-rolled orchestration, no agent framework, with a FastAPI backend and a React + TypeScript UI that answer through the same code path the eval harness measures.

failure-mode thinking

Measured on real 20-paper arXiv corpora, not the toy set: that's where hybrid RRF turned out to lose to dense alone (MRR 0.79 vs 0.94), which the 4-paper set had hidden. The evals also caught retrieval making the system more confidently wrong on unanswerable questions — 100% citation coverage isn't knowing when to abstain — and that's written up as an open gap, not buried.

  • Python
  • BM25 · dense · RRF
  • Gemini embeddings
  • Claim verification
  • Recall@K · MRR evals
  • FastAPI
  • React + TS
github ↗

PRJ—02

open source

SxQLear

database intelligence & archaeology tool

Born from a real problem: dropped into a 107-table production database with no documentation as a research assistant. SxQLear crawls and semantically maps unfamiliar SQL databases — a multi-signal inference engine (structural, lexical, statistical, semantic) proposes join relationships with confidence scores, and schema memory accumulates per project.

failure-mode thinking

Local-first by design: no cloud dependency, no data leaves the machine. Confidence scores instead of confident guesses — the tool says how sure it is, and asks a human before it believes itself.

  • Python
  • FastAPI
  • React
  • Embeddings
  • Schema inference
  • SQLAlchemy
github ↗

PRJ—03

1st / 450+ teams

Asclepius

clinical intelligence engine · Hack-n-Win 3.0, CGC University · with Team Deranked

Multimodal pipeline for brain-tumour progression analysis: a Temporal Cognition Engine runs delta analysis across longitudinal MRI scans, and a RAG layer (MongoDB Atlas + Gemini 2.5 Flash) generates therapy reports. Ingestion is automated with multi-agent validation. Took first place out of 450+ teams — and second for best use of MongoDB.

failure-mode thinking

Built to fail soft: if the vector layer fails, the system falls back to preserved deterministic paths instead of returning nothing — in a medical context, silence is harm.

  • PyTorch
  • MongoDB Atlas
  • Gemini 2.5 Flash
  • RAG
  • Multi-agent
github ↗

PRJ—04

1st runner-up

Paeon AI

compliance-first digital medical representative · Plaksha MediThon 2026

Real-time, source-verified drug intelligence for pharma reps and clinicians at the point of care: structured retrieval over CDSCO and FDA datasets, a Twilio voice interface, and a hybrid pipeline — deterministic retrieval first, Gemini as fallback. Built in 24 hours; first runner-up.

failure-mode thinking

Designed around what it must never say: guardrails block diagnostic or unsafe outputs entirely, and every answer is auditable back to its source.

  • FastAPI
  • React
  • Vercel
  • Twilio
  • Gemini 2.0 Flash
  • Guardrails
github ↗

clubs & community — where the hackathons get built

guilds joined.

LEADfeb '26 — now

Geek Room Plaksha

Technical Lead

  • Organising events, hackathons and the college fest.
  • Prayas — inter-university tech fest, technical lead.
  • Colossus — intra-college tech fest.
  • Trackshift '26 — built the website for Mphasis' Trackshift.
see the trackshift site ↗

AIfeb '26 — now

LEAP — AI@Plaksha

Technical Member

  • AI research on OLM (Open Language Model) — an open-source PyTorch library for training language models.
  • Contributed to OLM's LLM gallery: the ready-made model presets (GPT-2, Llama, Qwen, Phi, Gemma and more).
  • Organised events on campus.
openlanguagemodel ↗

SECaug '25 — now

Cyber Defense Club

Technical Member · Plaksha University

  • DaVinci Node at Colossus — an AI-jailbreaking event.
  • Capture the Flag — ran workshops on CTF techniques and cybersecurity ideals, twice.
  • Organising events, workshops and fests.

GDGdec '25 — now

Google Developer Group Ranchi

Community Member

  • Part of the local developer ecosystem back home — workshops and meetups on AI, ML and backend development.

jack of all trades, master of none…

…but oftentimes better than master of one.

off the clock.

▸ the shelf

  • The AlchemistPaulo Coelho — follow the omens; the road is the treasure.
  • The Subtle Art of Not Giving a F*ckMark Manson — choose what's worth caring about.
  • To Kill a MockingbirdHarper Lee — courage is doing right when you know you'll lose.

…and the shelf keeps going

  • The God Equation — Michio Kaku ♥
  • Diary of a Wimpy Kid — the whole series, still my favourite
  • It Ends With Us · It Starts With Us · Verity — Colleen Hoover

difficult roads truly lead to beautiful places.

— spiti valley, jul '26 · drag the ring, click a frame

viva la vida.

tools I reach for

moves learned.

CORE

  • Python
  • C++
  • SQL
  • Git

AI / ML

  • PyTorch
  • Scikit-learn · XGBoost
  • RAG pipelines
  • Guardrails · embeddings

BACKEND

  • FastAPI
  • Docker
  • JWT (OAuth2)
  • SQLAlchemy · Alembic

DATA

  • PostgreSQL
  • MongoDB Atlas (Vector)
  • SQLite
  • pandas · statsmodels

badges earned

  • Systems Design — high-level architecture
  • Building AI Agents with MongoDB
  • Microsoft: Fundamentals of AI Security
  • Data Fundamentals
  • Artificial Intelligence Fundamentals

if your system needs to keep working when things go wrong —

let's build something that survives.

direct routes

githubSw4ppleneel↗ linkedinSwapneel Premchand↗ instagram@swapneelfr↗