Devender Gollapally
Principal Software Engineer — I build production AI systems in regulated finance.
I ship agentic, human-in-the-loop LLM systems that operate under real regulatory constraints (SEC / FINRA), and I'm a co-author of ERC-1450 — now a Final Ethereum standard for putting real securities on a public blockchain. 26 years building the systems other engineers build on top of — from payments and securities infrastructure to the AI layer running on top of it today.
🌐 devender.me · 💼 LinkedIn · 📍 Los Angeles · Remote
🤖 What I'm focused on now
- Agentic AI in production — multi-step LLM workflows with humans in the loop, built for domains where a wrong answer has compliance consequences, not just a bad demo.
- Compliance-gated automation — AI that drafts, a person who approves; every action auditable. Built inside a regulated broker-dealer, with compliance as a design partner rather than a review gate.
- Evaluation & reliability — the unglamorous half of applied AI: making agent output measurable, idempotent, and safe to put in front of regulated workflows. I publish the numbers, including the ones that don't flatter my own claims.
- MCP tool design — what actually makes a tool surface usable by a model, measured rather than asserted.
🔗 Selected public work
| Project | What it is |
|---|---|
| ERC-1450 — RTA-Controlled Security Token Standard | A Final Ethereum standard (July 2026): the on-chain token defers to a registered transfer agent, so a real security can live on a public chain without breaking securities law. Co-author — I revived it from Stagnant through Draft → Review → Last Call → Final. Halborn-audited reference implementation (Solidity, UUPS, multi-sig), in production behind hundreds of deployed securities. |
| mcpwright — MCP servers for public data | Four published, typed, CI-gated MCP servers: SEC EDGAR, US Census, IRS SOI, FRED — plus mcpwright-core, the shared runtime they're all built on. On PyPI, in the official MCP Registry, and packaged as one-click Claude Desktop extensions. |
| mcp-tool-surface-eval | An eval harness for MCP tool-surface design decisions. Two experiments so far — one of which refuted my own hypothesis. I published that too. |
| ethers-v6-kms-signer | An ethers.js v6 signer backed by AWS KMS — sign EVM transactions without the private key ever leaving the HSM. |
| iofinnet-openapi-spec | Community-maintained OpenAPI 3.0 spec for the IOFinNet MPC custody API. |
✍️ Writing
I write about tool design, measurement, and building AI you can actually trust at devender.me:
- What Makes a Good MCP Tool Surface for an LLM — six principles, each footnoted to a real shipped server.
- When a Tool Description Is the Difference Between Right and Wrong — the follow-up eval: one caveat sentence took a model from 0% to 100% on correctly reading a capped statistic.
🏗️ Background
Before AI, I spent six years building the core of a securities platform that has moved $1B+ in compliant offerings: payments, escrow, secondary-market trading, the recommendation engine, the data warehouse, and the shared libraries every microservice is built on. Then I built the tokenization stack — hybrid AWS KMS + IOFinnet MPC custody, hundreds of securities live on Polygon — and the AI platform running on top of all of it.
🛠️ Tech
Python · FastAPI · TypeScript / Node.js · Java / Spring · Solidity / Hardhat
PostgreSQL · Kafka · Redis · AWS (ECS, KMS, RDS, SES) · Terraform
Claude · OpenAI · MCP · LangChain
🎒 Beyond the terminal
- ✍️ I write & self-publish children's books as D.G Sunny — the Red Robot series and The Cat Monster of Blackpine Mountain (my words and stories; illustrations with AI).
- 🎶 I play ukulele on YouTube (@gdevender).
Always happy to talk about applied AI in regulated domains, agent evaluation, or on-chain securities. Reach me via devender.me.