Available for consulting engagements

Systems from 0→1.
Owned end to end.

I'm Divyansh, a full-stack engineer and multi-time founder. I take systems from first commit to production and own them across frontend, backend, infrastructure, and the agent layer. Most recently I was sole architect and tech lead of SCM, an AI-agent governance platform that makes autonomous agents safe to run against FERPA-regulated student data.

View my work Work with me

$8M+/yr

Revenue locked in at Union Pacific

89%

Revenue-loss trajectory reversed

0→1

AI governance platform, sole architect

200+

Daily active users in 4 months at VortexeAI

Selected Work

Projects

Production systems where I owned the architecture, the trade-offs, and the outcome, from AI-agent governance to compliance-critical rail infrastructure.

◆

Sole Architect · Tech Lead

SCM — AI Agent Governance Platform

Autonomous agents, made safe for FERPA-regulated student data

The governance layer that lets AI agents touch student records safely. Every agent action resolves to the authority of the human it acts for; every autonomous workflow runs on a single-use, time-boxed capability grant; and student names are structurally absent from everything the LLM sees. Built as a Python/Temporal agent runtime with human-in-the-loop approval gates, a NestJS capability registry, and a 71-tool MCP integration, shipped to a live production pilot with a large customer.

Python / FastAPITemporal LangGraphNestJS + Prisma MCPFERPA Architecture

Read the architecture deep-dive

Mission Control — the SCM teacher dashboard showing live class proficiency goals, Claude-generated small groups, and per-student pacing

◉

Co-founder & CEO

VortexeAI

Low-code AI workflow automation, 0→1 in four months

A low-code, AI-driven workflow editor for product and engineering teams, taken from first commit to 200+ daily active users and a TechStars invitation in four months. The core is a deterministic multi-agent orchestration engine, graph traversal plus dependency resolution, running at 99.9% execution reliability, with a React/Next.js node editor and an LLM tool runtime with memory routing. 50+ automation patterns deployed; onboarding 40% faster.

Next.jsMulti-agent orchestration LLM runtimeLow-code

Read the write-up Visit the live site

The SPX 3D equipment configurator running across MacBook, MacBook Air, and iMac

◭

Co-founder & CTO

AetherisVis

Real-time 3D product configurators in the browser

A browser-based 3D equipment configurator for SPX, a luxury gym design brand, shipped as a commercial client product in one week. Real-time 4K/8K customization stays responsive under load through lazy loading and interaction-loop tuning. Built on a modular React + TypeScript visualization framework of 35+ reusable components that cut UI prototyping time 60% across client projects.

React + TypeScript3D / WebGL Real-time renderingClient delivery

Read the write-up See it live at SPX

✚

Lead Engineer

LUV Health

Self-pay healthcare marketplace with AI-assisted booking

A platform that helps patients book same-day appointments with hard-to-find specialists at transparent self-pay rates. I led end-to-end development of the full MVP: smart search with tailored relevance signals, practice profiles, doctor-facing pricing and visibility tools, and automated scheduling. A voice-based AI agent (Vapi) books appointments on the patient's behalf, cutting booking time from minutes to seconds and keeping the path from search to booked appointment as short as possible.

FastAPINext.js Voice AI (Vapi)Marketplace

Iso-Scapes title screen: a procedurally generated isometric low-poly landscape of forests, cliffs, and water, with a Re-Generate button

⬡

Core Architecture

Isoscapes

Procedural isometric worlds, re-generated at the push of a button

A procedurally generated, explorable 3D low-poly landscape built in Unity at UCSC. I built the core architecture: Perlin noise and falloff-map generation over a grid system, a Cell abstract data structure holding each block's attributes, the isometric camera with movement and zoom, and auto-tiled terrain driven by noise-based height values, prototyped in p5.js and ported to C#. Shipped as builds for both Windows and Mac.

Unity / C#Procedural Generation Perlin NoiseCross-platform

See it on the games page

Tracewell pipeline schematic — a flagged injection event passing through deterministic feature extraction into a single narrating LLM call

⌁

Solo Developer

Tracewell

Guardrailed insights for browser-agent trajectories

An insight pipeline for AI-agent run logs: deterministic Python computes every fact, one LLM call narrates them, and three independent layers stop prompt injections scraped off a web page from steering the report. Ships with an append-only run ledger and a live prompt-robustness eval harness.

FastAPIPydantic LLM GuardrailsReact

Read the write-up Visit the live site

Keyed-cache schematic — responses routed to their query keys, with a stale response structurally unable to land

▤

Case Study

UoP Library

Frontend architecture as an executive-grade case study

Took a half-working prototype and structurally eliminated an unplanted fetch race with TanStack Query, built optimistic cross-route cache coherence, and designed a Bloomberg-style tearsheet grid, documented in 20 logged decisions and a 134-finding self-audit.

React 18TanStack Query TypeScriptDesign Systems

Read the case study

Rail telemetry schematic — wayside, AEI, and CAD signals joined on trainID, with a flagged bearing on the rail line

▣

Technical Lead

Freezer-Car Modernization

Deterministic correlation over rail sensor networks

Reversed an 89% revenue-loss trajectory at Union Pacific and locked in $8M+/year. Built a deterministic correlation engine joining wayside, AEI, and CAD signals over Kafka, with ML-derived health thresholds surfacing derailment-risk bearings for crew intervention.

KafkaC++ ML ThresholdsRail Systems

Component tree schematic with metadata badges that make the UI legible to AI agents

◈

Solo Developer

Sense UI

AI-comprehensible component architecture

A metadata-driven component architecture that lets AI agents comprehend and operate UI, measured as a 40% lift in accurate agent interactions and 30% less integration debugging. Schema extraction and introspection pipelines deployed across 15+ projects.

ReactSchema Introspection Agent UX

Visit the landing page

▥

Solo Developer

Atlas Admin Panel

Sub-second search at 10,000+ concurrent users

An admin dashboard with sub-second prefix and fuzzy search via tuned Elasticsearch indexing. Role-based invites, cursor-based pagination, and Redis caching cut data-retrieval latency 35% on a zero-flicker Next.js + React 18 frontend.

NestJSElasticsearch RedisNext.js

Frontend source Backend source

▲

Founder & CTO

Overwhelming Studios

Two shipped UE5 titles, recognized by PlayStation Studios

An independent game studio where I shipped two Unreal Engine 5 titles solo: Packet Patrol, a CRT-styled network-defense game, and Vegan Zombie: Crop Apocalypse, an outback survival game with a procedural recipe system and parabolic damage falloff. Cloud auth and cross-device sync for 1,500+ players with 30% faster loads and 25% better retention. The work drew direct recognition from PlayStation's Head of UI/UX and reached promotion-support stage with PlayStation Studios.

Unreal Engine 5Gameplay Systems Cloud Auth + SyncGame Design

Explore the studio & devlogs

Vegan Zombie Crop Apocalypse title screen, set on an Australian outback road

Interactive

Run the architecture

Most portfolios describe systems. This one runs one: SCM's grant lifecycle, reimplemented in your browser. Mint a single-use permission slip and watch the whole pipeline execute.

Every idea from the SCM deep-dive is in this demo. The workflow never holds a credential; it holds a grant ID. Scope is read from the grant row, so the tampered classroomId=999 in the request gets ignored. Students reach the model as opaque refs (S1…S6), never as names. The LLM gets exactly one call. And every step appends to an audit ledger that nothing can edit.

When the run finishes, try Replay expired grant. The request fails with a 410 and a non-retryable error, because retries cannot un-expire a grant. A system is defined by what it refuses to do.

All client-side, no tracking, ~150 lines of vanilla JS. View source; that's what it's there for.

Experience

Resume

Full-stack engineer and multi-time founder. Known for shipping fast and defending every architectural decision from first principles.

Teaching Lab Studio

Mar 2026 — May 2026

Software Engineer III · Tech Lead, SCM

Sole architect and tech lead of SCM, taking it from 0→1 to a live production pilot with a large customer while owning engineering, product strategy, and forward-deployed customer work in parallel.
Re-architected the sensitive-data boundary for FERPA-regulated data from in-flight scrubbing to structural isolation: student PII never enters the AI region at all.
Built the capability registry, skill launcher, and per-invocation authorization model end-to-end (NestJS, Prisma, Postgres, Next.js, React) with time-bound grants and 7 FERPA-mapped audit events.
Built SCM into an identity provider for connected AI agents on a Python/FastAPI + LangGraph runtime: single-use grants, per-failure-class retry policies, ≤2 LLM calls per workflow.
Engineered the Temporal runtime with human-in-the-loop approval gates and deterministic, replayable workflows with re-validation closures.