8000 himanshu634 (Himanshu Mendapara) Β· GitHub
[go: up one dir, main page]

Skip to content
View himanshu634's full-sized avatar
🎯
Focusing
🎯
Focusing

Organizations

@nakartha

Block or report himanshu634

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
himanshu634/README.md

Hi, I'm Himanshu πŸ‘‹

I build AI agents and the software around them. This is where I experiment: document AI over financial filings, graph memory for agents, and agents that keep running long after you close the tab.

πŸ“ˆ Maester

Maester is my main project: an open-source portfolio manager for long-term investors, India-first and self-hostable. It watches your holdings, filings and prices, and tells you which decision is due, with the evidence behind it.

One rule runs through it: the model reads, code calculates, and every number cites its source.

What's under the hood:

  • Document AI over annual reports. Docling and RapidOCR turn report PDFs into searchable chunks. XBRL filings are parsed alongside, and each filing's PDF and XBRL stay paired.
  • LangGraph extraction. Finds the statement pages, fans out one branch per statement with Send, then loops extract β†’ deterministic arithmetic check β†’ targeted correction. Retries, a model-call budget, and bad corrections get rejected. Claude with structured outputs.
  • Hybrid RAG. pgvector HNSW over halfvec(1024) plus Postgres full-text search, fused with reciprocal rank fusion. A gate refuses queries when the embedding service serves a different model than the one the index was built with.
  • Provenance on every figure. The filing, page, unit, period and reporting basis. Missing data shows as unknown, never as zero.
  • Evals. maester-eval runs offline with dev and held-out splits. On held-out banks: 1,521 statement-sum checks across 65 filings, none failed.
  • MCP server. Seven typed, read-only tools, so AI assistants answer from your filings with cited figures.
  • Data. A 4,675-file BSE filings corpus (13 GB of PDFs and XBRL), plus read-only Zerodha holdings.

Next: named, versioned calculations (a margin is a calculation you can inspect, never a model's arithmetic), then return tracking, then autonomous monitoring that suggests and explains but never trades.

TypeScript, Hono, Drizzle, Zod, SvelteKit, Python, LangGraph, Postgres + pgvector. AGPL, pre-alpha, built in public.

πŸ€– AI agents and other experiments

The thread through all of it: agents you can trust with real data.

  • Citadel: a company brain on FalkorDB. Graph RAG with multi-agent memory, 22 reviewed Cypher tools, and a grounding check that rejects any citation the agent never saw. When the graph can't answer, a Keeper agent researches the gap and writes the fix back. Built in 3 days for Graph Hacks (FalkorDB Γ— WeMakeDevs); passes 119 of 120 graded eval runs.
  • Insider: no official source publishes live commodity prices. They move through commodity traders' WhatsApp groups, as scattered Gujarati and English messages. Insider reads those groups and turns them into clean, comparable mandi prices and a daily brief, so a buyer can see the price gap between mandis and act on the arbitrage first.
  • Claude Code is how I build: skill libraries, reviewer subagents, hooks. Side quest: Tank, a macOS menu bar app that moves Claude Code sessions between accounts (private for now).

πŸ› οΈ What I go deep on

  • Long-running agents in production. I run agents as durable jobs: LangGraph runs checkpointed in Postgres, so a run survives a crash and resumes where it stopped. Human approval gates in the middle of a run. Background jobs over a queue with idempotency keys and live progress over SSE. Retries, model-call budgets and timeouts. Agents never write straight to the system of record. Maester's autonomous monitoring is designed around the same pattern.
  • Graph databases. FalkorDB and Cypher in Citadel: a graph per company, reviewed read-only queries with timeouts, and graph RAG whose answers cite the nodes they came from. I use a graph where the question is about relationships that a vector index can't follow.
  • Vectors. pgvector HNSW over half-precision embeddings, hybrid search that fuses vector and full-text rankings, and an index that refuses queries from any model other than the one that built it.

πŸ™ GitHub

Open source: I wrote the Trino connector for WrenAI and wren-engine, and presented it on the Trino Community Broadcast (2024).

Himanshu's GitHub stats Top languages

πŸ“« Say hi

builtbyhimanshu.com Β· LinkedIn Β· X Β· Email

Pinned Loading

  1. maester maester Public

    Open-source portfolio management platform that watches your holdings, filings and prices, and tells you which decision is due, with the evidence behind it. India-first, self-hostable.

    TypeScript

  2. Insider Insider Public

    Turns trader WhatsApp groups into clean, comparable mandi prices and a daily market brief for commodity buyers. Read-only collector, bilingual (Gujarati/English), TypeScript monorepo.

    TypeScript

  3. citadel citadel Public

    A company brain with falkordb

    Python

  4. Herald Herald Public

    Manage your brand's mail.

    Python

0