HI·I'M MOKSHITH·KARIMNAGAR, IN

21, founding engineer at flashback labs. typescript, go, and most of the AWS bill.

442K assets · 590K tweets · 13d runtime

02 · ABOUT

HI, I'M MOKSHITH

I'M

Founding engineer at Flashback Labs. TypeScript, Go, and most of the AWS bill. If you're building something hard, let's talk.

The plumbing that decides whether a system actually holdsunder load — or just demos well.

— WHAT I CARE ABOUT

CURRENTLY BUILDING ›▌
Mokshith Rao

◆ KARIMNAGAR, IN · b. 2004

NEXT — SELECTED WORK

→ 03

03 · SELECTED WORK · 4 PROJECTS

2024 — 2026

THINGS I'VE
ACTUALLY

Production pipelines, multi-agent platforms, browser-native runtimes, on-device ML. Each card has the source — click through for the details.

AURA · MULTI-AGENT

PLANNER → 5 WORKERS → EVALUATOR

№ 01

AURA

Agentic Unified Reasoning Assistant.

A multi-agent AI workflow platform. Type a natural-language instruction — "draft an email to my team and add a calendar invite" — and a Planner → 5 Workers → Evaluator pipeline executes it across Gmail, Drive, Docs, Sheets, and Calendar autonomously.

  • Hierarchical Planner → 5 Workers → Evaluator with resumable DAG execution
  • AES-256-GCM encrypted OAuth tokens with 5-min pre-expiry auto-refresh
  • Realtime progress streaming via Supabase Realtime — full execution observability
  • Next.js 16
  • Gemini 2.5
  • Supabase
  • Google APIs
  • TypeScript
→ REPO

GROK PIPELINE · DEDUP

442K · 60 GB · 99.99%

№ 02

GROK PIPELINE

Distributed media ingestion at production scale.

A three-stage distributed pipeline — TypeScript scraper → Go image processor → enricher — ingesting AI-generated media across 10 rotating accounts. Triple-tier deduplication, deterministic stealth fingerprinting, ML-based face detection inline.

  • 442K deduplicated images (60 GB) at 99.99% S3 ↔ DynamoDB consistency in 13 days
  • Triple dedup: URL → SHA-256 → perceptual hash (Hamming ≤ 10)
  • SCRFD ONNX face detection in Go with goroutine pools across 27K+ hourly windows
  • TypeScript
  • Go
  • AWS SQS / DynamoDB / S3
  • Playwright
  • ONNX Runtime
→ PRIVATE REPO

CHEFMATE · ON-DEVICE ML

TFLITE · WEB + ANDROID

№ 03

CHEFMATE

AI recipe and meal-planning, web + Android.

AI-powered recipe generator and meal planner shipping as both a Vercel-hosted web app and a native Android app via Capacitor. On-device ingredient detection, calorie/macro tracking, 7-day auto-fill meal planning, subscription tiers.

  • On-device ingredient detection via TFLite (TensorFlow.js + custom Vite plugin)
  • Mifflin-St Jeor BMR → TDEE nutritional model with macro-validation per slot
  • PKCE OAuth2 with Android deep-link callbacks; subscription-tier quota enforcement
  • React
  • Capacitor
  • Supabase
  • Gemini
  • TensorFlow.js
  • Vite
→ REPO

AUTODEV · WEBCONTAINER

NL → MULTI-FILE PROJECT

№ 04

AUTODEV

Browser-native AI code editor.

An in-browser AI code editor that converts natural language into multi-file projects, running Node.js dev servers in-page via WebContainers. Async agent pipeline with multi-LLM routing. Honorable Mention at hackathon.

  • Async agent pipeline (Inngest) with multi-LLM routing and structured tool-call error handling
  • Live file-tree and editor sync via Convex; Node.js dev servers running in-page (WebContainers)
  • GitHub OAuth + Octokit for repository import / export
  • Next.js
  • WebContainers
  • Inngest
  • Convex
  • Multi-LLM
  • Clerk
→ REPO

NEXT — PRODUCTION METRICS

→ 04

04 · TELEMETRY · IN PRODUCTION

FILE 04 / METRICS

[ TELEMETRY · SEP 2024 — APR 2026 ]

PRODUCTION
METRICS.

What the work actually moved. Pulled from the Grok pipeline run log and the 48-hour hiring platform sprint — real counters, not vanity rounding.

K

IMAGES PROCESSED

K

TWEETS INGESTED

GB

S3 THROUGHPUT

%

DEDUP CONSISTENCY

DAYS

CONTINUOUS RUNTIME

HOURS

HIRING PLATFORM SHIPPED

END FILE 04 / NEXT — STACK

6 SIGNALS · LIVE

05 · STACK · 6 TIERS · 25 TOOLS

WORKBENCH

TOOLS I
REACH

Not tutorials I've finished. Things in active use across production projects. Click any chip to see what it actually does in the runtime.

● LIVE TRACE · A SAMPLE REQUEST · HOVER A LAYER TO PAUSE

01

CLIENT / RUNTIME

RUNTIME

  • Node.js
  • Browser / WebContainers
  • Goroutines
  • Android / Capacitor
02

APP / FRAMEWORK

APP / FRAMEWORK

  • Next.js
  • React
  • Vite
  • Express
  • Tailwind
03

INFRA

INFRA

  • AWS SQS
  • Inngest
  • Vercel
04

DATA

DATA

  • Postgres / Supabase
  • DynamoDB
  • Convex
  • S3
05

AI / ML

AI / ML

  • OpenAI
  • Anthropic
  • Groq
  • Gemini
  • ONNX / SCRFD
  • TFLite
06

LANGUAGES — SUBSTRATE

LANGUAGES

  • TypeScript
  • Go
  • SQL

NEXT — NOW

→ 06

06 · NOW · UPDATED 27 APRIL 2026

SNAPSHOT

WHAT I'M
DOING

Inspired by /now pages. A snapshot, not a stream. I bump it on edit, not by clock.

STATUS

ONLINE · —

FOCUS

GROK PIPELINE PHASE-2

LOCATION

KARIMNAGAR · UTC+5:30

UPTIME

—

SINCE FOUNDING-ENGINEER START · FLASHBACK LABS

CURRENT BUILD

Phase-2 grok pipeline · 1M asset target

ROTATES · 1 / 3

Phase-2 of the Grok pipeline — pushing toward a 1M-asset target with tighter face-detection latency on the processor pods.

Deeper Go concurrency patterns and ONNX model surgery; reading the Tigerbeetle internals.

ChefMate v1 — meal planner across web + native Android with on-device ingredient detection.

Final-semester B.Tech CSE at Jyothishmathi Institute. Graduating April 2026.

OPEN TO FOUNDING-ENGINEER / INFRA / AI-INFRA ROLES

KARIMNAGAR · UTC +5:30

NEXT — CONTACT

→ 07

07 · CONTACT

LET'S TALK

SOMETHING.

→mokshithrao1481@gmail.com
GITHUBLINKEDINRESUME

MOKSHITH RAO · KARIMNAGAR, IN · 2026

BUILT WITH NEXT.JS · GSAP · MOTION · DEPLOYED ON VERCEL

↑ BACK TO TOP