AI & Machine Learning
- LangGraph
- Google ADK
- Gemini
- OpenAI
- Claude
- Groq
- Hugging Face
- LiteLLM
- RAG
- Vector search
- Embeddings
- Prompt Engineering
- Structured outputs
- Tool calling
- AI agents
- LLM failover
- Isolation Forest
- Z-score
- IQR
- STL
- ML pipelines
Available for opportunities
Hi, I'm Jayant Singh
|
- 0+
- Years Exp
- 0
- Companies
- 0+
- Projects
- 0+
- Technologies
JS
01 About
End-to-end ownership, from data to deployed product.
I'm an engineer who likes the whole path: data modelling, backend architecture, AI orchestration, and the frontend that finally makes it useful. That range is the point — I can take a problem from "learners aren't getting what they need" all the way to a product someone opens every morning.
Right now I'm building an AI-powered learning platform for German tax-exam preparation at ShortHills AI — turning structured tax-domain content into personalized learning paths, with learner onboarding feeding downstream personalization and Neo Chat, an AI assistant that supports learners in the context of the lesson or quiz they're on. FastAPI, Next.js, PostgreSQL/pgvector and LangGraph.
Before that I designed the AI Data Integrity & Root-Cause Analysis Platform: a 4-algorithm anomaly detection ensemble, an autonomous Google ADK + Gemini agent with 10 tools, and a 6-step pipeline that replaced work people used to do by hand. And before AI took over my time, years of backend systems in PHP, Laravel and .NET — architecture, standards, and the unglamorous reliability work.
On the side I build Praxis924, a live AI learning & mentor platform I designed, wrote and deployed alone — RAG-grounded lesson generation on pgvector, a deterministic Mentor OS behind an LLM mentor, and real-time messaging, running on AWS behind my own CI/CD. Plus Sukoon, an AI financial decision assistant.
I'm based in Noida, India, and I default to automation over repetition.
◈
End-to-end ownership
Design → deploy, no handoff gaps
◆
AI architecture
Systems & integration, not demos
▲
Automation-first
If it repeats, it gets scripted
●
Backend + ML
The combination is the edge
02 Skills
The toolkit.
Grouped by what I actually reach for, not by how long the list can get.
Backend
- Python 3.14
- FastAPI
- SQLAlchemy
- Pydantic
- Alembic
- REST APIs
- WebSockets
- SSE streaming
- Redis
- DBOS
- Caching
- Connection pooling
- Pub/sub
- MVC
- Microservices
- Background jobs
- pytest
Frontend
- Next.js 15
- React 19
- TypeScript
- Tailwind
- Zustand
- Monaco
- Angular
- JavaScript
- HTML5
- API integration
- UX design
Cloud & DevOps
- AWS EC2
- AWS RDS
- AWS Lambda
- Secrets Manager
- Docker
- Docker Compose
- Nginx
- Let's Encrypt
- GitHub Actions
- GitLab
- CI/CD
- LAMP stack
Data
- PostgreSQL
- pgvector
- MySQL
- AWS RDS
- Redis
- Playwright
- BeautifulSoup
- Web scraping
- Data validation
- Chunking & embedding
Also in the belt
- C#
- ASP.NET
- PHP
- Laravel
- CodeIgniter
- MVC
- OOP patterns
- TDD
- Unit testing
- Git
- Drupal
- WordPress
- Security
03 Experience
Where I've done the work.
-
2024 — Present Current
Software Engineer 2 / ShortHills AI
AI Product Engineer & Backend Engineer
German Tax Exam Preparation Platform
- Building an AI-powered learning platform for German tax-exam preparation, converting structured tax-domain content into personalized learning experiences.
- Developed learner onboarding capturing goals, interests, background and preferences, feeding downstream personalization.
- Built personalized learning paths and dashboards that adapt curriculum recommendations to each learner profile.
- Built Neo Chat, an AI learning assistant reachable directly or contextually from lessons and quizzes, grounded in learner and lesson context.
- Working across FastAPI, Next.js, PostgreSQL, pgvector and LangGraph on learner and AI-assistance workflows.
AI Data Integrity & Root-Cause Analysis Platform · Automobile
- Built a data-quality system for an automobile analytics platform serving 18,000+ dealers, validating the revenue, profit and performance figures two production applications (J2 and J3) reported to dealers and dealer groups.
- Designed a production-grade anomaly detection system using a 4-algorithm ensemble (Isolation Forest, Z-score, IQR, STL) over dealer data cleaned upstream by the data science team.
- Built root-cause analysis by reverse computation — backtracking a flagged figure through the formulas applied during cleaning to reconstruct the original input and pinpoint where the value actually broke.
- Automated a 6-step pipeline that replaced a manual two-engineer comparison: Playwright scrapes J2 and J3, cross-compares dealer data, runs anomaly detection and RCA, then scores each dealer’s data quality.
- Developed an autonomous AI agent (Google ADK + Gemini) with 10 tools for unattended root-cause investigation, plus dashboards and an SSE-streaming AI chat interface.
- Implemented thread-safe caching, SQLAlchemy connection pooling, background jobs, and CI/CD with GitHub Actions on AWS.
FastAPIPythonNext.jsLangGraphGoogle ADKGeminiPostgreSQLpgvectorAWSDocker
-
08/2018 — 04/2021
Software Developer / Chetu India Pvt. Ltd.
Noida, India · Team Lead
- Started with PHP, WordPress and Drupal, then contributed to a large insurance-domain platform built with Laravel, MySQL and a web frontend.
- Led the migration of the insurance platform's backend from PHP/Laravel to Python.
- Progressed to Team Lead, managing 7–8 developers across Python, PHP, Angular, WordPress and Drupal.
- Handled client communication, requirement gathering, technical planning and architecture across multiple projects.
- Contributed to AWS Lambda migration, API integrations, CQRS-based system design, unit testing and production deployments.
PHPLaravelPythonMySQLAngularCQRSAWS Lambda
-
04/2018 — 08/2018
Web Developer / Trivialworks Solutions
Lucknow, India
- Built a full crowdfunding web application from scratch using Laravel and CodeIgniter.
- Integrated PayPal and Authorize.net payment gateways for secure transactions.
- Implemented a subscription module with PayPal for recurring donations.
- Integrated Google Location and Kroger APIs for store search and product donations.
- Managed end-to-end development across cross-functional teams.
LaravelCodeIgniterPayPalAuthorize.net
-
04/2017 — 04/2018
PHP Developer / Pamac Finserve Pvt. Ltd.
Pune, India
- Worked on a support project using CodeIgniter, MySQL, JavaScript, and jQuery.
- Resolved assigned tickets, maintaining both code and database.
- Built sample projects with CodeIgniter, Laravel, and Core PHP; used Git for version control.
CodeIgniterMySQLjQueryGit
04 Projects
Selected work.
The systems I've designed and shipped. Open any card for the full breakdown.
Live Solo project
Jun 2026 — PresentPraxis924 — AI Learning & Mentor Platform
A full-stack AI learning platform I designed, wrote and deployed alone — RAG-grounded lesson generation, a deterministic Mentor OS behind an LLM mentor, and real-time messaging, running on AWS behind my own CI/CD.
FastAPIPython 3.14PostgreSQLpgvectorRedisDBOSNext.js 15React 19TypeScriptDockerAWS EC2Nginx
A look at it
How the platform works
Admins define a Technology → Framework → Lesson syllabus, and an LLM pipeline auto-generates each lesson — four structured sections plus graded exercises — with a self-review loop that re-scores and improves any section falling below threshold.
How it works
- RAG grounding layer — datasets are chunked, embedded via Hugging Face and stored in pgvector; retrieval uses cosine similarity with keyword and first-N fallbacks, so generation stays grounded even when vector search misses.
- Multi-provider LLM failover — Groq → Gemini → Hugging Face with automatic failover, Redis caching and deterministic offline fallbacks. A single provider outage never breaks generation, chat, grading or interviews.
- Mentor OS — deterministic learner intelligence and decision-making kept separate from the LLM-powered mentor experience, so recommendations are engine-derived and explainable rather than improvised by a model.
- Durable generation — DBOS Postgres-backed workflows mean a mid-pipeline failure leaves a lesson safely re-runnable instead of half-written, plus a token-frugal "fill gaps" path (0 LLM calls when complete).
- Real-time messaging — WebSockets over a Redis pub/sub backplane for cross-worker fan-out: presence, typing indicators, read receipts, heartbeat reconnects. Messages persist even when live delivery is down.
Also shipped
Lucy, the SSE-streaming AI mentor interface that delivers engine-derived recommendations and guidance, grounded in the lesson you opened it from · AI-graded coding challenges · stateless mock interviews · a Next.js 15 App Router frontend with a Monaco practice workspace and admin console · Docker Compose on AWS EC2 behind Nginx + Let's Encrypt · GitHub Actions CI with Alembic migration-drift detection gating an SSH-based CD pipeline · 25+ pytest suites covering generation, RAG, messaging and AI-safety behaviour.
★ Flagship ShortHills AI · Automobile
2024–25AI Data Integrity & Root-Cause Analysis Platform
A QA system for the analytics behind 18,000+ automobile dealers. It validates what two production applications report, detects anomalies in the numbers, traces each one back to its root cause, and scores every dealer’s data — work that was previously manual, or simply not done.
FastAPIPythonGoogle ADKGeminiPlaywrightSQLAlchemyIsolation ForestSTLSSE streamingAWSDocker
The problem
Dealer data arrived as Excel exports. A data science team filtered, normalized and applied formulas to clean it, and the backend then served the result to two customer-facing applications — J2 (the legacy platform) and J3 (its replacement) — where dealers and dealer groups saw their performance, revenue, profit and loss. Nobody could answer the obvious question: is any of this actually correct? Verification meant two engineers pulling numbers out of both applications by hand and comparing them in a separate tool.
What I built
- Anomaly detection ensemble — Isolation Forest, Z-score, IQR and STL run together over the dealer figures. Each catches what the others miss, and the four signals are combined into one verdict, which is what keeps false positives low enough for people to trust it.
- Root-cause analysis by reverse computation — the hard part. Rather than only flagging a bad number, the engine backtracks it through the transformations the data science team applied, reconstructing the original input to pinpoint where the value broke. The output is a cause, not an alert.
- 6-step automated pipeline — I folded the two engineers’ manual comparison into the platform. One run now drives a Playwright script that scrapes J2, scrapes J3 for a single dealer or many, cross-compares the two, runs anomaly detection, performs RCA, and scores the result — unattended.
- Dealer data-quality scoring — every dealer comes out with a score plus the specifics: what is wrong, what is missing, and which figure to look at first.
- Autonomous AI agent — a Google ADK + Gemini agent with 10 tools that investigates an anomaly on its own and reports back, replacing an engineer reading dashboards.
- Dashboards and AI chat — a purpose-built dashboard for the findings, with an SSE-streaming chat interface for asking the system about what it just found.
Engineering
FastAPI backend with thread-safe caching and SQLAlchemy connection pooling · Playwright-driven ingestion across both applications · real-time SSE streaming APIs · background jobs running the pipeline on its own cadence · CI/CD via GitHub Actions with AWS-based environment configuration.
★ Current ShortHills AI
2026German Tax Exam Preparation Platform
An AI-powered learning platform that converts structured German tax-domain content into personalized learning experiences — adaptive learning paths, dashboards, and an AI assistant that answers in the context of the lesson you are on.
FastAPINext.jsPostgreSQLpgvectorLangGraphPython
What I built
- Learner onboarding — captures goals, interests, background and preferences, and feeds those signals into downstream personalization rather than parking them in a profile page nobody reads.
- Personalized learning paths — dashboards and curriculum recommendations that adapt dynamically to each individual learner profile.
- Neo Chat — an AI learning assistant reachable directly or contextually from a lesson or quiz, answering with the relevant learner and lesson context in hand.
- Platform work — FastAPI, Next.js, PostgreSQL, pgvector and LangGraph across the learner and AI-assistance workflows.
Solo project · Fintech AI
2026Sukoon — AI Financial Decision Assistant
A mobile-first PWA that helps people understand income, commitments, spending capacity and savings opportunities — built around decisions, not generic financial trivia.
AIPWAPersonalizationRecommendations
What I built
- Decision-oriented recommendations — the system answers "what should I do about this?" rather than restating a balance the user can already see.
- Daily Brief and Weekly Review — personalized recurring experiences built from the user's financial context and past decisions.
- Mobile-first PWA — designed for the phone first, using a modern web and AI application architecture.
Fintech · Trivialworks
2018Crowdfunding Platform
Built from scratch: campaigns, recurring donations via PayPal subscriptions, and Authorize.net checkout, with store search and product donations.
LaravelCodeIgniterPayPalAuthorize.net
What I built
- Full application from scratch — a complete crowdfunding web app on Laravel and CodeIgniter.
- Dual payment gateways — PayPal and Authorize.net integrated for secure checkout.
- Recurring donations — a subscription module built on PayPal for repeat giving.
- Product donations — Google Location and Kroger API integration so donors could find stores and donate goods, not only money.
- End-to-end ownership — managed delivery across cross-functional teams.
Architecture · Chetu
2019–21Insurance Platform — Laravel → Python Migration
A large insurance-domain platform on Laravel and MySQL. I led the migration of its backend from PHP/Laravel to Python while a team of 7–8 kept the product shipping.
LaravelPythonMySQLCQRSTeam lead
What the work involved
- Backend migration — led the move from PHP/Laravel to Python on a platform that could not simply be switched off for the transition.
- Team leadership — managed 7–8 developers across Python, PHP, Angular, WordPress and Drupal.
- Client-facing architecture — requirement gathering, technical planning and architecture across multiple projects.
- System design — CQRS-based design, API integrations, unit testing and production deployments.
Cloud · Chetu
2020Legacy → AWS Lambda Migration
Moved legacy systems onto AWS Lambda, improving performance and scalability while keeping the business running through the transition.
AWS LambdaPHPServerless
What the work involved
- Serverless migration — legacy workloads moved onto AWS Lambda for better performance and scalability.
- No downtime for the business — the platform kept serving customers throughout the move.
- Deployment discipline — unit testing and production deployment practices carried over with the code.
No projects in this category.
05 Education
Foundations.
2017
Bachelor of Technology — Computer Science & Engineering
Krishna Group of Institutions — Kanpur, UP
Coursework in C, ASP.NET, HTML, CSS.
2013
Intermediate — Science
Triveni Kashi Inter College — Unnao, UP
2011
HSC — Science
Triveni Shivmaul Singh K.P Singh Memorial Inter College — Raebareli, UP
06 Contact
Let's build something.
Open to conversations about AI systems, backend architecture, and problems worth automating.
Hiring? Grab the full CV.
Complete experience, skills, and education — PDF, 173 KB.