Solomon Mithrasolomon-mithra.dev

Full-Stack AI Engineer | TypeScript / Python | 3 years

I build full-stack AI products with public receipts.

Currently Founding Engineer at Budgeting Parent, building a 0->1 personal-finance platform for early earners and continue to contribute to open source.

Email me
PUBLIC RECEIPTSUPDATED FROM PUBLIC APIS
✓Gateia npm downloads1,084 downloads↗✓stylelint PR #8643 mergedmerged; 11.5k stars↗✓mem01Long-term memory for AI agents↗
✓M.S. Computer Science, University of the PacificGPA 3.9 | Phi Kappa Phi
01

Case studies

FLAGSHIP · OPEN SOURCE · JAN 2026 - PRESENT

Gateia

The deterministic verification layer for enterprise AI

The problem. LLM output is probabilistic; production systems are not. Teams ship AI features that can leak secrets, emit unsafe markup, or break the data contract downstream code depends on, and "another LLM checks it" is not a guarantee.

The approach. Gateia is an open-source TypeScript SDK that sits between the model and the user as a deterministic gate. Every output must satisfy a Zod contract and code-based policies, using regex and logic engines instead of AI judging AI. It ships with markup-safe, secrets-safe, pii-safe, and finance-safe policies, custom policies typed against the output contract, and audit-mode logging with per-decision trace IDs.

The proof. Live on npm since January 2026, with eight published versions, 1,000+ downloads, and an MIT license. The chat on this page dogfoods it: every answer passes through gateia.verify() before rendering, and each response carries its trace ID. The design reasoning is written up in the two essays below.

Budgeting Parent

A live AI money copilot that never invents the numbers

The role. Founding Engineer since March 2026, owning the full stack of a live personal-finance product: frontend, API route handlers, data, and the AI layer.

The architecture. Two halves, deliberately separated. A deterministic, unit-tested cash-flow engine computes every number: safe-to-spend, safe-to-save, monthly burn. The Claude copilot sits on top purely to explain them: an agentic tool-use loop streamed to the browser over Server-Sent Events with live per-tool status updates. The model never invents a figure.

The infrastructure. Durable Plaid bank sync (link-token exchange, incremental cursor-based transaction sync, webhooks), per-user rate limiting with semaphore-based concurrency control, token-cost controls, and a freemium billing gate on Clerk + Supabase, backed by 140 automated tests.

The proof. Live at budgetingparent.com. The same thesis that drives Gateia — deterministic logic owns the truth, the model only explains it — running as a real product.

Next.js 16 · React 19 · Supabase · Clerk · Plaid · Claude (SSE streaming)

Live site ↗

mem01

Long-term memory for AI agents

The problem. Append-only agent memory rots: contradictions and stale facts pile up, and every conversation turn pays for them in tokens.

The approach. mem01 stores memory as beliefs with an explicit lifecycle — add, update, supersede, invalidate, merge — and serves token-budgeted, conflict-filtered recall with zero LLM calls on the read path. Self-hosted: a FastAPI service over PostgreSQL + pgvector, packaged with Docker Compose.

The proof. The source is public on GitHub with a tested write/read pipeline and a runnable Docker Compose stack.

Python · FastAPI · PostgreSQL · pgvector · Docker

Live site ↗Source ↗
Budgeting Parent0->1 personal-finance platform for early earners; full-stack architecture plus an AI decision engine over financial inputs.↗RelevectLocal context engine for AI agents with semantic search, PDF/Markdown/text re-indexing, pgvector, and a FastAPI interface.↗stylelint contributionMerged PR #8643 fixed false positives in function-linear-gradient-no-nonstandard-direction for CSS variable names containing direction substrings.↗LiteLLM contributionPR #31942 fixes the proxy's projected-spend budget alerts to respect daily/weekly budget reset windows instead of always projecting to calendar month-end, with regression tests.↗more on GitHub — 45 public repos↗
02

Experience

Mar 2026 - Present

Founding Engineer — Budgeting Parent (budgetingparent.com)

  • Built a streaming AI money copilot live in production: a Claude agentic tool-use loop streamed over Server-Sent Events with live per-tool status updates.
  • Designed the deterministic cash-flow engine beneath the copilot: pure, unit-tested computation of safe-to-spend, safe-to-save, and monthly burn. The LLM explains the numbers and never invents them.
  • Built durable Plaid bank-sync infrastructure: link-token exchange, incremental cursor-based transaction sync, and webhook handling.
  • Hardened the API for production with per-user rate limiting, semaphore-based concurrency control, input validation, token-cost controls, and a freemium billing gate (Clerk + Supabase), backed by 140 automated tests.

Oct 2025 - Jun 2026

Software Engineer — Cher (real-estate technology startup)

  • Built an AI loan agent in the borrower flow: a multi-iteration Claude tool-use loop over the live loan database, with read-only tools, iteration caps, and a prompt-cached system + tools prefix to cut latency and token cost.
  • Built CRM and loan-workflow systems, document upload/validation flows (Next.js + FastAPI), and REST APIs with PostgreSQL and Cloudflare R2.

Mar 2023 - Aug 2023

Software Development Engineer — Kennect Inc.

  • Modernized a Vue 2 monolith into a Nuxt 3 micro-frontend, cutting bundle size by 25%.
  • Built reusable UI components on REST APIs, increasing feature adoption by 18%.

Aug 2021 - Mar 2023

Software Engineer - Automation (RPA) — Wipro

  • Built automation processing 5,000+ daily financial transactions.
  • Cut error rate from 11% to under 3% and processing time by 35%.

Aug 2023 - May 2025

M.S. Computer Science — University of the Pacific

GPA 3.9/4.00 · Phi Kappa Phi Honor Society (top 10%)

03

Writing

Jan 31, 2026

Treat AI Output as Untrusted Input

A short argument for treating model output like any other untrusted system boundary.

#ai #infrastructure #security

Jan 30, 2026

Probability Is a Liability in Production

Why production AI systems need deterministic checks around probabilistic model output.

#ai #typescript #opensource

04

Toolbox

Languages

TypeScript · JavaScript · Python · SQL

AI / LLM

AI agents / tool-use loops · SSE streaming · prompt engineering · prompt caching · RAG · embeddings · vector search · pgvector · evals · LLM verification (Gateia)

Frontend

React · Next.js · Vue · Nuxt · Hono · HTML · CSS · responsive web

Backend

Node.js · FastAPI · REST APIs · microservices · API design · Flask

Cloud & Infra

AWS EC2 · AWS S3 · Cloudflare Workers · Cloudflare R2 · Docker · Kubernetes · CI/CD · Vercel · Inngest

Databases

PostgreSQL · Redis · MongoDB · Supabase · Firebase

Also

Agile/Scrum · unit testing · Blue Prism RPA · Git · Figma

05 CONTACT

Hiring for AI engineering? Let's talk.

Email is fastest — or ask the chat below for a summary of anything on this page first.

solomonmithra99@gmail.comLinkedIn ↗GitHub ↗