Selected work

03 / 05

Kepler AI

Infrastructure for long-horizon bioinformatics agents. Each conversation gets its own Linux VM, a persistent workspace, and state that survives a crash.

Role
AI Engineer Intern
Timeline
May – Aug 2025
Company
Kepler AI (Keplogic) · PearX S25
Stack
Python, Firecracker, Firestore, S3, Pydantic
AGENTSBIOINFORMATICS
The problem

A chat box is not enough for a bioinformatics workflow.

Researchers at biotech customers wanted to run real analyses through natural language, over private enterprise data, in sessions that ran for hours. That means the agent needs a computer to run code on, memory that outlives a single request, and a way to recover when something dies mid-run.

Isolated execution

One Linux VM and one workspace per conversation.

I built Kepler's sandbox execution infrastructure on Firecracker. Every conversation is provisioned an isolated Linux VM and a persistent S3 workspace, so agents can safely run code and browser-based bioinformatics workflows without touching anyone else's data.

Persistence and recovery

State that survives a crash, and streams that don't flood the database.

Cross-session persistence lives in Firestore. Streamed tokens are batched before they are written, which cut database writes by about 95% without losing conversation state.

Real-time interrupts and atomic locks mean a conversation can be stopped, resumed, or recovered after a crash without two workers fighting over the same state.

The tool surface

About ten tools with one schema, one retry policy, one result shape.

I built roughly ten agent tools across code execution, browsing, and file access. Each uses a standardized Pydantic schema, the same retry behavior, and the same result handling, so adding a tool did not mean re-teaching the agent how to read output.

What shipped
  • Agent infrastructure for long-horizon agents handling private enterprise data, with cross-session persistence in Firestore, batched token writes, real-time interrupts, and atomic locks.
  • Sandbox execution infrastructure on Firecracker with an isolated Linux VM and persistent S3 workspace per conversation.
  • About ten agent tools across code execution, browsing, and file access with standardized Pydantic schemas, retries, and result handling.
Outcomes
Database writes
~95% fewerBy batching streamed tokens.
Agent tools
~10Code, browser, and file access.
Isolation
1 VM / conversationFirecracker with a persistent S3 workspace.
Recovery
Interrupts + locksSafe resume after crashes.

This page covers the 2025 bioinformatics-agent product I worked on. Kepler has since moved into verifiable AI for finance; none of that work is mine.

Next project · 04 / 05 AsyncHealth