I build software that lives close to the metal — simulation kernels, mixed-signal verification, operating systems, and ML — and I'm fascinated by how AI can make low-level software faster, smarter, and easier to debug.
I'm a software engineer at Cadence Design Systems, but the short version of me is simpler: I like building things that run close to the hardware, and I like making them fast.
I got here from an unusual direction. I started in web and machine learning — Django backends, computer-vision models for medical imaging, an AI writing assistant — then kept drilling downward until I was writing a microkernel from scratch, a Verilog simulator, and a deterministic dataflow OS for fun. The lower the layer, the more interested I get.
Right now that curiosity lives in mixed-signal verification R&D at Cadence, on the simulation engine behind Xcelium — profiling hot paths in modern C++, reworking the event-scheduling regions where analog and digital meet, the kind of work where a 10% speedup on a real customer workload is a good week. It's deep, underexplored, and exactly the sort of problem I want to spend years on.
The thread tying it all together: I'm convinced AI is about to change how we build low-level systems — automated root-cause analysis, agentic debugging, ML that spots anomalies in code most models never see. I'm already building those workflows at work, and chasing the same idea across my side projects.
The places where I want to spend the next several years — and where my side projects keep pointing.
Kernels, schedulers, simulators, lock-free data structures, zero-copy I/O. Profiling with perf/Valgrind and tuning the inner loop until the numbers actually move.
Agentic workflows that triage crashes, do automated root-cause analysis, and reason about code most models never see — letting AI meet bare metal.
Event-driven simulation and the analog↔digital boundary at Cadence. The newest territory for me, and the one I most want to leave a mark on.
From web internships in India to mixed-signal R&D in Silicon Valley.
Heaviest at the bottom of the stack, comfortable all the way up.
A curated cut — the projects I'm actually proud of, from bare-metal kernels to AI systems. Filter by domain.
A from-scratch bare-metal AArch64 microkernel for the Raspberry Pi 4 — boots without Linux into an EL1 shell. ~5K LOC of C + assembly: real-time scheduler, VFS + FAT32, a hand-written TCP/IP stack with lock-free RX/TX rings, and zero-copy DMA networking.
A deterministic OS for low-jitter sensor → inference pipelines: zero-copy buffer pools, SPSC ring queues, p50/p95/p99 latency tracking, and explicit backpressure. Built in phases with RFCs, sanitizers, warnings-as-errors, and CI — engineering rigor as a feature.
A modern-C++ Verilog simulator implementing the same event-wheel model as production RTL tools — Active / Inactive / NBA / Postponed regions per SystemVerilog LRM §4.5 — plus a peephole event-coalescing optimizer. The direct conceptual ancestor of my Cadence work.
A long/short US-equity bot blending unusual math: graph-Laplacian diffusion residual alpha with a persistent-homology (TDA) regime gate. Bias-safe walk-forward sim, live IBKR data, commission/slippage/borrow modeling, and a Streamlit diagnostics dashboard.
An AI voice agent that answers a clinic's landline and books appointments in Hindi / English / Marathi. Multi-layer: a Raspberry-Pi edge box tapping the phone line, LiveKit voice agents, Sarvam Indic STT/TTS, Claude for reasoning, FastAPI backend — with DPDP / ART-Act compliance baked in.
Detects cervical-spine fractures from CT scans using YOLOv8 over DICOM imagery, wrapped in a Django front-end with a full detection pipeline. An automated second read for assessing trauma severity.
Implementation of "Learning Auxiliary Monocular Contexts Helps Monocular 3D Object Detection" — predicting 3D bounding boxes from a single camera, built on mmdetection / mmdetection3d.
A deep-learning malware detector using a binary-representational image approach — turning executables into images for a CNN to classify. Shipped as a Dockerized Django web app, and the basis of my published paper.
A personal investment dashboard that unifies broker accounts — Plaid and ICICI integrations on a Next.js 14 + Supabase stack — into one view of where the money actually is.
A complete hotel/lodge operations system — bookings, guests, rooms, housekeeping, POS, inventory, invoicing, payments, expenses and shifts — across a stack of Django apps, containerized with Docker Compose.
A sentiment-analysis engine for Amazon product reviews, built as a full software-engineering exercise — DOI, CI/CD, code of conduct, the works. Scrapes, cleans, and scores customer feedback at scale.
A no-friction budget processor: ingest bank-statement CSVs, auto-categorize into rent / utilities / groceries, and export clean, Numbers-friendly summaries. Small, single-file, and genuinely used.
…and ~50 more across the years on github.com/blackhawk005
I think best when I'm moving. The ideas that end up in my code usually show up somewhere on this list first.
Long rides where the problem I was stuck on quietly solves itself around mile 15.
Give me an open highway and a playlist and I'll happily lose a whole weekend to it.
Same loop as engineering — tweak the inputs, taste the output, iterate until it's right.
Progressive overload for the body. Consistency compounds the same way good code does.
Whether it's mixed-signal simulation, systems work, or teaching AI to understand low-level code — if it's hard and interesting, I want to hear about it.