# Harsh Sinha

> Prev @ Multibagg AI · National Finalist IFF-FinTech Olympiad’24 · IITP'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 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).
- 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.
- 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 Technology, [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

### 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*

A read-only research dashboard where LLMs paper-trade the Nifty 50 under realistic constraints, with rankings, trade history, portfolio analytics and deterministic 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

*Structured event intelligence for Indian equities*

A financial event-intelligence platform that organizes filings, disclosures and news into structured, traceable events with materiality and source context for Indian-equity research.

[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

- [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, React, Next.js, Node.js, Fastify, Tailwind CSS, OpenAI, PostgreSQL, Prisma, Redis, BullMQ, Pinecone, Qdrant, Docker, Azure, Grafana, Git, GitHub, Razorpay, FFmpeg, Vitest, Zod