pok8dev

Mumbai · --:-- IST · B.Tech AIML

Samuel D'Souza

I ship the messy middle — agents that call tools, watches that update over the air, models that try to see an attack before it lands.

currently forecasting packets

  • 0projects
  • 0repos
  • 0PRISM F1%
  • 0internships done
scroll the bench

about

Not a stack tourist. I stay until it boots on the real board.

Most of my work lives in the gap between a working demo and a production constraint: 4 MB of flash, a 60-student classroom, a false-positive rate a SOC won't tolerate. I treat those limits as the brief, not a later problem.

Right now that's PRISM with Aryan & Shaun — transformer world models on network traffic, on a cloud box that stays up. Previously: firmware internship at Mozek Healthtech (completed Aug 2026).

work

Internships, and the server that doesn't sleep.

I run my own cloud boxes so long jobs don't have to sit on a laptop. Google Cloud first, then Oracle Cloud when I needed something that stays up.

  • Always-on VMs for labs that have to run overnight — Docker, dashboards, the boring glue.
  • DNS and a reverse proxy in front of pok8.dev. Private access for anything that shouldn't be public.

constraint · always-on

A box that stays up. GCP, then Oracle. DNS in front. Nothing here is a how-to.

Wearable health-tech in Boisar. Internship completed August 2026. The brief was getting over-the-air updates onto a production wearable that was already tight on flash, then making sure the rest of the fleet could actually receive them.

  • Shipped OTA firmware and merged it into the production baseline.
  • Built an OTA server (Node/Express/WebSocket) for check, download, and device updates.
  • Added a multi-level sleep state machine to ease idle power strain.
  • UART and hardware bring-up on the bench — flash budget and boot behavior as first-class constraints.

constraint · production wearable · OTA

No spare flash, no spare power. The work was reducing strain on both until OTA could ship and stay shipped.

Trained and shipped a custom YOLOv9 detector for disposal bins — annotation through training through a published artifact. Not a pretrained demo with a new README.

Garbage-Bin-Detector →

Goregaon. Front-line hardware, software, and web issues at a company that actually ships sites. Learned how to talk to a stuck human before talking to a stuck machine.

lab

Things I actually built.

PRISM lab console — live multi-model attack forecast session

security · flagship

PRISM

Temporal transformer world models that forecast network attacks K-steps ahead. Dual telemetry (NetFlow + Scapy), 85% F1 vs 54% logistic baseline, false positives cut from ~41% to 16.7%. 222 MITRE techniques. Docker attacker/target loop that retrains on evasions.

PyTorch · Scapy · Docker · React · FastAPI

  • PRISM attack forecasting
  • Pixie desktop avatar
  • InsightForge datathon canvas
  • ACAMS classroom vision
  • AIScanner offline legal OCR
  • Behavioral CAPTCHA mouse biometrics
  • gpt_functions… pre-tools agent
  • Labshare2 lab lending
  • OTA server device updates
  • Always-on box Oracle + GCP
  • Spectacular Spectacles Pi HUD glasses

stack

The bench, not a keyword dump.

talks to models

Python, PyTorch, Lightning, Groq, Gemini, OpenRouter, llama.cpp, FAISS, PPO

talks to silicon

C/C++, ESP-IDF, Arduino, NimBLE, UART, OTA partitions, Raspberry Pi, ADB

talks to people

FastAPI, Next.js, React, TypeScript, Node, WebSockets, PySide6, Docker, Caddy, Cloudflare, GitHub Pages

talks to the rack

GCP, Oracle Cloud, Cloudflare, Linux

school

B.Tech AIML, SJCEM · 2025–28 · diploma at Pravin Patil Polytechnic · Datathon / Hackverse / Hackstomp

ping

If you want the walkthrough, not the pitch deck.

[email protected]
GitHub LinkedIn Mumbai