# Harsh Sinha

> Prev @ Multibagg AI · National Finalist IFF-FinTech Olympiad’24 · IIT Patna'27 · Working on AI Agents, Quant and Backend

AI Engineer · Prev Founder's Office @ Multibagg AI · National Finalist IFF-FinTech Olympiad '24 · IIT Patna '27.

Source: [https://www.harshsinha.dev](https://www.harshsinha.dev). Machine-readable index: [https://www.harshsinha.dev/llms.txt](https://www.harshsinha.dev/llms.txt). JSON: [https://www.harshsinha.dev/api/about](https://www.harshsinha.dev/api/about).

## About

- I am currently looking for AI Engineering roles around AI agents, quant, and backend. Previously, I was Founder's Office & AI Engineer at [Multibagg AI](https://www.multibagg.ai).
- I'm in my final year of undergrad at [IIT Patna](https://www.iitp.ac.in/). Was also the national finalist at [IFF–FinTech Olympiad '24](https://www.linkedin.com/posts/harshsinha12_fintecholympiad2024-fintech-nationalfinalist-activity-7259242419387314176-kO12), among the top 30 out of >1 lakh candidates.
- I love building AI agents — for finance, payments, data pipelines, news, Instagram analysis, and most workflows I can automate.
- I've mainly worked on [Ask Iris](https://www.multibagg.ai/ask-iris) and [Multibagg AI](https://www.multibagg.ai), which has answered over 500K+ user queries and helps investors daily. It's loved by users and the sharks on [Shark Tank India Season 5](https://www.linkedin.com/posts/shark-tank-india_namitathapar-sharktankindia-sharktankindiaseason5-ugcPost-7418286077645312000-fXBB).
- Fun fact: got into finance pre-COVID, watching Dad invest in the stock market. Investing since 2019 — was not 18 yet, lol 😅 — generally profitable, with a few F&O losses too. [Nine out of ten people lose in F&O](https://www.sebi.gov.in/reports-and-statistics/research/jan-2023/study-analysis-of-profit-and-loss-of-individual-traders-dealing-in-equity-fando-segment_67525.html) — stay away unless you actually know what you're doing.

## Contact

- Email: sinha.harshsep@gmail.com
- Twitter: https://x.com/sinhaharsh12
- LinkedIn: https://www.linkedin.com/in/harshsinha12/
- GitHub: https://www.github.com/harshsinha-12
- Résumé: https://drive.google.com/file/d/1Iq1ZV_sMimkoGrNui8gR6_VP04-el2TD/view?usp=sharing

## Experience

### Founder's Office & AI Engineer — [Multibagg AI](https://www.multibagg.ai)

Jan 2025 - Jun 2026

Stack: Next.js, Node.js, Python, OpenAI, Prisma, PostgreSQL, Pinecone, Qdrant, Redis, BullMQ, Grafana, Azure, Docker

  - Worked directly with the founder to build and scale production AI systems across agent orchestration, retrieval, evaluation, financial data and user-facing product workflows.
  - Built core workflows for [Ask Iris](https://www.multibagg.ai/ask-iris), a multi-agent investment research assistant that answered 500K+ user queries ([Iris launch](https://www.linkedin.com/posts/biased-human_today-we-are-launching-the-most-powerful-ugcPost-7398653952919101440-vvJ_)). Orchestrated specialized agents and tools for SQL, RAG, web search, citations and streaming across 100K+ documents and 20M+ records.
  - Created evaluation harnesses across stock, portfolio, screener, ETF, index and industry agents using custom test sets, LLM-as-a-judge scoring, tool-call and citation validation, latency tracking and failure analysis.
  - Built production document-intelligence pipelines for IPO RHPs, ETF factsheets, annual reports, investor presentations and earnings-call transcripts using Docling/OCR, typed schemas, queue workers and structured extraction. Optimized retrieval across [Pinecone](https://www.pinecone.io) and [Qdrant](https://qdrant.tech) with hybrid search, re-ranking, metadata filters and page-level citations, reducing vector infrastructure costs by up to 80%.
  - Developed an AI Screener Agent that translates natural-language investing queries into SQL and filter operations, improving reliability through schema mapping, evaluations, logging, guardrails and prompt optimization.
  - Built financial and real-time market automation across 300+ ratios and indicators and 6K+ companies, covering news, exchange announcements, transcripts, sentiment, market breadth and sector rotation. The resulting [automated X posts](https://x.com/sinhaharsh12/status/1975865353705320477) reached 3.2M+ impressions in six months.
  - Designed and tested Redis Cluster deployments across Docker and Azure VMs, validating primary-replica failover, key-access patterns, deployment behaviour and migration strategy.

## Education

### Bachelor of Science, [Indian Institute of Technology, Patna](https://www.iitp.ac.in/)

Aug 2023 - May 2027

- Major: Computer Science and Engineering
- Minor: Data Science and Artificial Intelligence

## Projects

### Tourist

*Runnable city foundation · cloud agent, memory, and learning in progress*

Imagine Clash of Clans for your codebase: every file is a village building, and you can watch builders work on it. I'm building a cloud agent to fix issues and open PRs, with six layers of memory and a reinforcement learning loop planned to make its decisions more accurate and cost-efficient.

[Live](https://web-sigma-orpin-64.vercel.app) · [GitHub](https://github.com/harshsinha-12/tourist)

Stack: TypeScript, Next.js, React, GitHub, Vitest, Python, OpenAI, Agentic Systems, LangChain, LangGraph, LangSmith, PostgreSQL, Redis, Pinecone, Qdrant, Reinforcement learning

### DevSweep

*Native SwiftUI · Trash-only cleanup, never rm -rf*

A native macOS utility that finds regeneratable developer files: node_modules, build caches, DerivedData, explains why each folder was detected, and moves only the items you approve to Trash.

[GitHub](https://github.com/harshsinha-12/dev-sweep)

Stack: Swift, SwiftUI, macOS

### Room FloorPlan

*Room AI take-home · geometry owns centimetres, the LLM cannot write wall lengths*

Imagine assessing a house for an insurance claim: every wall and patch of damage needs measuring. This local pipeline uses iPhone LiDAR to map the floor area and room geometry, then computer vision and LLMs to find and classify damage, producing a measured floor plan with evidence for an estimate.

[GitHub](https://github.com/harshsinha-12/roomplan)

Stack: Python, Swift, OpenAI, OpenCV, NumPy, LiDAR, Reinforcement learning

### Library Survey

*Replacement-cost survey · models propose, they don't write count or money*

An iPhone LiDAR app uses a live AI guide to walk a surveyor through the library's floor plan, damage, shelves, and other objects. Computer vision counts books from their spines; ISBN, barcode, and title evidence help identify editions; web searches draft local prices. After sealing, two models independently review the same evidence, and Jev suggests recapture or human review under fixed rules. An offline reinforcement-learning loop is designed to improve that routing, while a signed report estimates the whole library's replacement cost.

[GitHub](https://github.com/harshsinha-12/library-estimate)

Stack: Swift, FastAPI, Redis, OpenAI, Cloudflare, OpenCV, YOLO, LiDAR, Reinforcement learning

### RecoveryOS

*Razorpay AI Buildathon · verified Test Mode recovery flow*

An explainable revenue-recovery system for Razorpay merchants: AI proposes one bounded action, deterministic policy guards execution, and durable workflows follow failed payments to auditable outcomes.

[Live](https://rzpy-agent-web.vercel.app) · [GitHub](https://github.com/harshsinha-12/rzpy-agent)

Stack: Next.js, Fastify, Redis, PostgreSQL, OpenAI, Razorpay

### LLM Trading Arena

*Frontend · Nifty 50 paper-trading arena*

An experiment where LLMs paper-trade Nifty 50 stocks using market news, including each story's sentiment and importance, custom technical indicators, quantitative analysis, and portfolio strategies. The read-only dashboard lets you follow their trades, rankings, and portfolio performance, with reproducible Redis-backed replay.

[Live](https://the-llm-trading-arena-frontend.vercel.app) · [GitHub](https://github.com/harshsinha-12/-the-llm-trading-arena-frontend)

Stack: Next.js, TypeScript, Redis, OpenAI, Tailwind CSS

### Vritta AI

*Ranked financial news · traceable events · Indian equities first*

A financial-news platform for following topics and regions on demand. It fetches news, filings, and disclosures, ranks their importance, and connects related items to structured, traceable events with source context and materiality. Its current focus is Indian equities, helping researchers move from a headline to the evidence behind it.

[Live](https://vritta-one.vercel.app/) · [GitHub](https://github.com/harshsinha-12/Vritta)

Stack: Next.js, TypeScript, Redis, Vitest, BullMQ, Pinecone, Azure, PostgreSQL

### Instagram Creative Intelligence

*Evidence-first analysis · reports, transcripts and frame sampling*

A multi-agent analysis pipeline that ranks public Instagram posts, extracts video and audio evidence, and turns measurable creative patterns into an adaptable strategy report.

[Live](https://instagram-analysis-red.vercel.app) · [GitHub](https://github.com/harshsinha-12/instagram-analysis)

Stack: Next.js, TypeScript, OpenAI, FFmpeg, Zod

### LLM Trading Arena Engine

*Backend · quantitative features and auditable simulation*

A TypeScript paper-trading engine for LLMs on the Nifty 50, with technical features, portfolio-aware risk context, Redis state and reproducible execution rules.

[GitHub](https://github.com/harshsinha-12/the-llm-trading-arena-backend)

Stack: TypeScript, Node.js, Redis, BullMQ, Quant Finance

### Go Rabbit

*Issue → validated patch → draft PR*

An agentic contributor assistant that scopes Go issues, scans repositories, generates and validates focused patches, and prepares draft pull requests behind explicit safety gates.

[Live](https://go-rabbit-sable.vercel.app) · [GitHub](https://github.com/harshsinha-12/go-rabbit)

Stack: Next.js, TypeScript, OpenAI, GitHub, Zod

## Articles

- [How I'm Building Tourist: A Cloud Coding Agent That Lives in a City — Markdown](https://www.harshsinha.dev/articles/building-tourist-a-cloud-coding-agent-in-a-city/article.md): Tourist's city foundation is running. Here is the build plan for the cloud coding agent, scoped memory, dynamic context, GitHub delivery, and learning that should follow. [Human-readable article](https://www.harshsinha.dev/articles/building-tourist-a-cloud-coding-agent-in-a-city).
- [Linear Regression: Where Did That Prediction Come From? — Markdown](https://www.harshsinha.dev/articles/linear-regression-a-line-that-learns/article.md): From a 20-minute delivery estimate to LLM latency: explore how a line learns, why we square its mistakes, and when a good-looking prediction stops being useful. [Human-readable article](https://www.harshsinha.dev/articles/linear-regression-a-line-that-learns).
- [GPT-6 Astra: When the Model Becomes the Operator — Markdown](https://www.harshsinha.dev/articles/gpt-6-astra-when-the-model-becomes-the-operator/article.md): GPT-6 Astra is not just a stronger reasoning model. Its computer-use gains, critical cyber capability, and monitorability limits change how we should design and supervise AI agents. [Human-readable article](https://www.harshsinha.dev/articles/gpt-6-astra-when-the-model-becomes-the-operator).
- [Building AI Agents That Know When to Stop — Markdown](https://www.harshsinha.dev/articles/building-ai-agents-that-know-when-to-stop/article.md): A practical design for bounded agent loops: finish when the work is good enough, compact before context degrades, and stop safely when progress stalls. [Human-readable article](https://www.harshsinha.dev/articles/building-ai-agents-that-know-when-to-stop).
- [Memory as the Missing Layer in Self-Improving AI Agents — Markdown](https://www.harshsinha.dev/articles/self-improving-agents-memory-missing-layer/article.md): A practical architecture for agents that turn experience into safer, measurable improvements through memory, sandboxed experimentation, evaluation, and replay. [Human-readable article](https://www.harshsinha.dev/articles/self-improving-agents-memory-missing-layer).

## Hackathons & certifications

- **IFF–FinTech Olympiad '24** — National Finalist. Top 30 of >1 lakh candidates at the India FinTech Forum olympiad (with IFTA). ([proof](https://www.linkedin.com/posts/harshsinha12_fintecholympiad2024-fintech-nationalfinalist-activity-7259242419387314176-kO12))
- **Mine The Model · Celesta IIT Patna** — 2nd Place. Stock-price ML contest by NJack ML IIT Patna & Cynaptics IIT Indore — beat the benchmark.
- **Summer of Quant 2024** — Certificate of Merit. 6-week Elementary & Advanced quant finance programme by Quant Club, IIT Kharagpur.
- **Complete DS, ML, DL & NLP Bootcamp** — Certificate of Completion. 101.5-hour Krish Naik bootcamp covering data science, ML, deep learning and NLP. ([proof](https://ude.my/UC-e70c868b-2859-46b3-92ab-a73e1aa25ade))
- **100xdevs · 0-100 Full Stack** — Certificate of Achievement. Completed Harkirat Singh's 0-100 Full Stack Web Development course (Jul 2024). ([proof](https://100xdevs.com))
- **Mathematics for Data Science & GenAI** — Certificate of Completion. 23-hour Krish Naik course — maths from basics to advanced for data science and GenAI. ([proof](https://ude.my/UC-72be0351-7746-4035-a500-aa11a65fb1f6))
- **JPMorgan Chase · Software Engineering** — Job Simulation. Forage sim: stock data feed, JPMorgan tools, trader visuals, and an open-source bonus. ([proof](https://www.theforage.com))
- **Python Data Structures · UMich** — Course Certificate. University of Michigan on Coursera — Python data structures (Feb 2024). ([proof](https://coursera.org/verify/ZLMC62M7TF3D))
- **Programming for Everybody · UMich** — Course Certificate. University of Michigan intro to Python on Coursera (Aug 2023). ([proof](https://www.coursera.org/account/accomplishments/verify/S55PZXYVJJWM))
- **Overview of Data Visualization** — Project Certificate. Coursera guided project on data visualization fundamentals (Aug 2023). ([proof](https://coursera.org/verify/DP73QZUJ39Z9))

## Tech stack

TypeScript, JavaScript, Python, Swift, React, Next.js, Node.js, Fastify, Tailwind CSS, OpenAI, LangChain, LangGraph, LangSmith, Agentic Systems, Reinforcement learning, PostgreSQL, Prisma, Redis, BullMQ, Pinecone, Qdrant, Docker, Azure, Grafana, Git, GitHub, Razorpay, FFmpeg, Vitest, Zod