§ 1 · THESIS

Electrical engineer in Atlanta. Neuromorphic hardware for ten years — memristors on a PCB, spikes on Loihi, neurons in VHDL — and lately a lot of agentic systems, off-hours, with an LLM in the loop.

This site is part working notebook, part workbench — research on the left, builds on the right. The notebook traces the through-line of neuromorphic work; the bench is what I’ve been building on evenings and weekends to learn how agentic systems actually behave.

Currently Research Engineer at GTRI — embedded ML, neuromorphic, computer vision. Looking for neuromorphic / mission-driven ML, or sales engineering at an AI company doing something worth caring about.

FIG.1 V_m vs. t · two action potentials, simulated

§ 2 · LINEAGE

Same instinct, different substrates.

Click a year for the long version.

§ 3 · EXPERIENCE

Where the time went.

CURRENTLY AT GTRI

2018 — now

CURRENT

Research Engineer · Georgia Tech Research Institute

2018 – present · Atlanta, GA

Started on Intel’s Loihi: deployed spiking neural networks via Kapoho Bay, wrote the host-side pipeline and core-mapping logic, built per-core spike-rate telemetry to diagnose the inevitable distribution mismatch between training and silicon. Authored leaky-integrate-and-fire neurons in VHDL for custom neuromorphic silicon.

Branched into broader embedded ML / CV on Jetson and similar edge platforms when programs needed it — sensor pipelines, on-device inference, the usual integration work that makes a research demo behave like a system. Co-authored two papers and a demo across the neuromorphic stretch.

Most recent stretch: more applied embedded ML and computer vision, plus quiet exploratory work on what LLM agents can and can’t do inside an experimental harness.

WORK

Loihi · Kapoho Bay · VHDL · embedded ML · CV

STACK

Intel NxSDK · embedded C · VHDL · Python · Jetson

OUTPUT

2 papers · 1 demo

2017 — 2018

Hardware Engineer · Rain Neuromorphics

2017 – 2018 · employee #5 · age 19

Joined as employee #5 at 19. Hand-assembled physical analog neuron circuits — memristive devices wired as LIF neurons on PCBs — and brought up the test rigs that drove them. The first time I held a working spike on a scope. Still the cleanest lesson I’ve had in how much physics a circuit will quietly do for you if you let it.

WORK

Memristive LIF neurons · PCB bring-up

STACK

Memristive HW · KiCad · scope-driven debug

TAG

first job · age 19

§ 4 · BENCH

The bench, off-hours.

2 LIVE · 1 SHIPPED · 1 WIP

These are mine, on my time. They’re where I learn agentic systems from the inside — not from a podcast. Nothing here was built at GTRI or with GTRI resources.

Multi-brain LLM-driven trading harness — fundamentals, technicals, news, options flow, sentiment, event/macro — running a paper portfolio in 14-day cycles. Confluence radar collapses six signal types into a single grade per ticker; a reasoning view exposes the prompt chain that produced each decision.

Currently running paper-only. The point is to study how the agents actually reason and disagree, not to claim a number.

STACK

Next · Recharts · Supabase · Anthropic SDK

SIGNALS

6 brains, 1 grade

MODE

paper portfolio

BUILD

Solo · Claude · 6 wks

→ open meridian readme

A private site for my book club — ten friends, a book a cycle, one of us picks each round. Replaces a sprawl of Google Calendar + Drive folder + group chat with one place: magic-link auth, schedule, books index, audiobook streaming behind signed URLs.

Stood up in a single sitting yesterday and shipped to the group the same night. The fun part wasn’t the code — it was watching how short the gap can be between “wouldn’t this be nice” and friends actually using it, when an LLM is doing the typing with you.

STACK

Next · Supabase · signed URLs

USERS

10 · book club

URL

papercuts.cc

BUILD

Solo · Claude · 1 evening

→ open papercuts

Two agents: a hunter that surfaces roles I’d care about (filters defense / generic gov / mission-less), and an applicant that drafts tailored cover letters per posting. The point is to learn agentic loops on something I’ll feel the cost of when they fail.

Currently dogfooding v0. Hunter is live; applicant is gated behind manual review until the false-positive rate stops being embarrassing.

STACK

Python · LLM · scrapers

STATUS

WIP · v0 dogfood

BUILD

Solo · Claude · ongoing

readme

Designed in conversation with Claude across two committed directions, then synthesized into one: notebook column on the left, workshop rail on the right. Built the way the rest of the bench is built — a long evening of iteration with an LLM in the loop, until the layout felt like the work it’s describing.

STACK

HTML · vanilla JS · taste

URL

you’re on it

BUILD

Solo · Claude · 4 days

§ 5 · CONTACT

Mail the lab.

vishalp@thak.io

Atlanta-based, remote-friendly. Open to neuromorphic / connectomics / mission-driven ML, or sales engineering at an AI company doing something I’d actually use.