Shwetank Shivhare — Building LucidLoom

About

Models in production, not slides in a deck.

I'm a technology leader and engineer based in Singapore. Right now I'm building LucidLoom AI: TinyTales, a school platform of eighteen microservices, and TutorBee, a homework tutor that teaches instead of answering.

The throughline of the last fifteen years is credit risk: probability of default, climate risk and stress testing for banks that move real capital. Eight of those years were at DBS Singapore, leading up to twelve engineers across Singapore, India and China, deploying 88 microservices and 200+ applications under MAS and Basel oversight.

My instinct is to break a hard problem into pieces small enough to reason about, then ship them, from requirements to production. Systems that hold up under audit, scale and scrutiny.

Roots in Pune and Chennai → building from Singapore

Core stack

Python · FastAPI · PySparkExpert

Java · Spring BootExpert

LLM agents · RAG · MCPExpert

ML: XGBoost · TensorFlow · InsightFaceAdvanced

Cloud: PCF · OpenShift · GCP · DockerAdvanced

Data: PostgreSQL · pgvector · Oracle · SparkStrong

Selected work

Three systems worth talking about.

2025→ now

LucidLoom AI: TinyTales and TutorBee

Founder & Principal Engineer · Singapore · Mar 2025 → now

A multi-tenant school platform and a consumer AI tutor, architected and built end to end by one person. FastAPI and Node.js services on PostgreSQL and Redis, behind Traefik with zero-trust tunnelling, self-hosted for data sovereignty. Eighteen microservices in production.

Face-recognition attendanceInsightFace embeddings on pgvector. Roll-call from a camera feed.

Constraint-solver timetablesOR-Tools CP-SAT with maker-checker approvals.

LLM

Socratic AI tutorClaude and DeepSeek routed by mode, with an anti-answer integrity classifier.

Content moderationML filter on school messaging; computer-vision guardrails on images.

Identity and billingPasswordless OTP, JWT RBAC with 18 roles, GST-compliant subscriptions.

MCP

Agent-ready platformAn MCP server for AI agents and an LLM-driven end-to-end test harness.

1,200+ functionalities lucidloom.ai ↗ tinytales.lucidloom.ai ↗ tutorbee.lucidloom.ai ↗ FastAPINode.jsPostgreSQLInsightFaceOR-ToolsTraefik

2020→ 2024

Credit Scoring Model ecosystem

DBS Bank, Singapore · VP, Tech Manager & AI/ML Lead

Architected and deployed 88 microservices generating regional PD, LGD, EAD, ICRR and ACRR ratings across Asia, spanning AIRB, SME, climate risk, stress testing and watchlist prediction. Led a twelve-person cross-regional team, built the risk feature mart and the model-monitoring control tower, and drove the migrations (Tibco → Spring Boot/Python, Oracle → MariaDB, Solaris → Linux) that cut cost by 40% and lifted performance by 30%.

88 microservicesPD / LGD / EADMAS · Basel II/IIIPySparkSpring Boot−40% cost

2016→ 2020

CRANE: default prediction engine

DBS via Helius Technologies · ML developer

The full lifecycle of an XGBoost model estimating the likelihood of a customer defaulting within twelve months: data, features, training, and the Spring Boot services that put predictions in front of the business. Sister project: Sticky Deposit, retention-pattern analysis on the deposits book.

XGBoost12-month horizonSpring BootFull lifecycle

The path

Fifteen years, one throughline: ship what's hard.

Mar 2025 → now

Founder & Principal Engineer

LucidLoom AI, Singapore

Solo-building TinyTales and TutorBee: eighteen microservices, face-recognition attendance, CP-SAT timetabling, an LLM tutor with a Bayesian mastery model, and the identity, billing and DevOps stack underneath them all.

Jun 2020 → Sep 2024

Vice President · Tech Manager & AI/ML Lead, Risk

DBS Bank, Singapore

Led twelve engineers across three countries delivering 200+ applications in credit-risk modelling and regulatory reporting. Built an internal LLM assistant for risk analysts that cut compliance-report turnaround by 35%, and the control tower that monitors every model in the portfolio.

Dec 2016 → May 2020

Senior Consultant (client: DBS)

Helius Technologies, Singapore

Built CRANE, an XGBoost default-prediction model with a twelve-month horizon. Contributed to the Sticky Deposit project and migrated legacy Tibco BE to Spring Boot, cutting runtime by 30%.

Sep 2015 → Nov 2016

Tech Lead (client: Western Union)

Opus Software Technologies, India

Led the migration of Tibco BW 5.x to 6.3 across core money-transfer systems and designed the Profile Service transaction logic in Java.

Jun 2014 → Sep 2015

IT Analyst (client: JetBlue)

Tata Consultancy Services, India

Managed 22 developers across the Customer, Crew and Flight domains. Delivered Auto Check-In and the Flight Notification System on Tibco and Java.

Feb 2011 → May 2014

System Engineer (client: British Telecom)

Tech Mahindra, India

Subject-matter expert for the WFMT Ethernet product. Built modules in Java, J2EE and Tibco; led design and code reviews across teams.

2010

B.E. Computer Science

RGPV University

Where it started. SCJP, OCJP and Oracle SQL certified along the way.

The AI ledger

This page was made by an AI. Here is the receipt.

Shwetank supplied a résumé, a photo and taste. Everything else, including the design, the copy, the poems, the code and all ten languages, was produced by Claude in a single conversation. The numbers below are counted live from this page.

Build ledger

theshwetank.com · index.html · v4

ModelClaude Fable 5.1 · Anthropic

SessionOne conversation · 3 September 2026

Output—

Languages—

Words translated—

Renditions—

Design systems—

A human team, roughly—

The AIone sitting

✦ generated · not typed

Where a human team's hours would go

Visual design, three systems

Front-end build and motion

Copywriting and structure

Translating ~1,000 words into nine languages

Nine creative renditions

Review across ten languages and three looks

A human team, roughly

—

The AI

one sitting

Estimates assume experienced professionals at a normal pace. The translations were checked by the same model that wrote them; a native reader may still find a phrase to improve, and the source is one file away.

How it was made

Read the old site, the résumé PDF and the photo.
Wrote the English copy once, as a dictionary of strings.
Translated that dictionary into nine languages, in parallel.
Designed three looks on one set of design tokens.
Rendered every language and look in a headless browser, and fixed what broke.

Building LucidLoom: AI for the schools your kids go to.

Models in production, not slides in a deck.

Core stack

Three systems worth talking about.

LucidLoom AI: TinyTales and TutorBee

Credit Scoring Model ecosystem

CRANE: default prediction engine

Fifteen years, one throughline: ship what's hard.

The same fifteen years, rewritten nine ways.

This page was made by an AI. Here is the receipt.

Build ledger

Where a human team's hours would go

How it was made

4 AM is a decision, not an accident.

Got a hard problem
worth shipping?

Models in production, not slides in a deck.

Core stack

Three systems worth talking about.

LucidLoom AI: TinyTales and TutorBee

Credit Scoring Model ecosystem

CRANE: default prediction engine

Fifteen years, one throughline: ship what's hard.

The same fifteen years, rewritten nine ways.

This page was made by an AI. Here is the receipt.

Build ledger

Where a human team's hours would go

How it was made

4 AM is a decision, not an accident.

Got a hard problemworth shipping?

Got a hard problem
worth shipping?