Bilal Khan.
Computer Engineering student at Cal Poly SLO.
- Latest
- Quantitative Analyst Intern, Helwa AI
- Previously
- Software Engineer Intern, Alpha Crest
- Focus
- Systems · Hardware · Quant research
- Graduating
- May 2027
About
“For the things we have to learn before we can do them, we learn by doing them.”
Selected projects
The four I'd show you first, then everything else. Hover to preview, click for the full write-up.
Featured
04- Built at Cal Hacks 13.0
Gives one goal to a whole team's Claude Code agents, keeps their files from colliding, and returns one tested branch. Wide goals fan out with map/reduce.
- Built with
- TypeScript · SpacetimeDB · MCP · Claude Code · React
- Result
- Verified on two computers: both machines' Claude workers built a goal's tasks in parallel, and the work merged into one branch with the tests run. 478 automated checks pass, without calling Claude.
A 32-bit RISC-V processor with interrupt support, built in SystemVerilog from the datapath up and run on a Basys 3 FPGA.
- Role
- Two-person team. We built the whole processor: the control unit's state machine and decoder, the ALU, register file, immediate and branch-target generation, memory, and the control and status registers that handle interrupts.
- Built with
- SystemVerilog · RISC-V RV32I · RISC-V assembly · Vivado · Basys 3 (Artix-7)
- Result
- Interrupts verified in simulation: the CPU finishes its current instruction, saves where it was, runs the handler, and returns. Then ran our own RISC-V assembly on the FPGA, driving the board's display.
- Pre-release
An Android app that records lectures, transcribes them on the phone, and makes them searchable by words or by meaning.
- Role
- Built it end to end: the Expo app with on-device Whisper, the FastAPI + MCP server with its own OAuth server, and the MongoDB Atlas data layer.
- Built with
- MongoDB Atlas · FastAPI · Expo / React Native · whisper.cpp · MCP
- Result
- Topic queries 140× faster (8.7 s → 62 ms over 1,000 lectures) after redesigning the data model; 185 API tests run against a live MongoDB.
- In development
My senior capstone: optimizing work for Tenstorrent's Blackhole p150, an AI accelerator card.
More projects
07- In development
Firmware for a device that samples motion continuously, detects significant events, and keeps the readings from before and after each one.
- Role
- One of a three-person team, building the firmware up one subsystem at a time on an STM32 Nucleo board.
- Built with
- Embedded C · STM32 (NUCLEO-L476RG) · STM32CubeIDE · HAL · LIS3DH accelerometer · I²C
- Result
- In progress: build, flash, debug and serial output are working, and accelerometer readings are being converted to engineering units. Now on timer-driven sampling.
- Research tool
A multi-timeframe range detector for TradingView with a Kalman-filtered trend guard and non-repainting breakout signals, tested against random entries on 15 years of data across three markets.
- Role
- Designed and built it end to end: the Pine Script indicator, a Python port that reproduces it bar for bar, and the validation harness.
- Built with
- Pine Script v6 · TradingView · Python · pandas · Kalman filtering
- Result
- Before costs, +0.10R per trade better than random entries matched on market, side, month, session and risk size (z = 7.7, about 11,600 trades over 15 years).
- In development
An unmanned ground vehicle that patrols a warehouse. Still being scoped.
Classifying Parkinson's from spiral and wave drawings — four architectures, tried in sequence until one held.
- Role
- Built and compared all four approaches, from hand-crafted HOG features to a dilated CNN.
- Built with
- Python · TensorFlow · Keras · scikit-learn · XGBoost
- Result
- Settled on a regularized dilated-convolution network, evaluated on confusion matrices rather than accuracy alone.
Predicting molecular solubility and toxicity by treating molecules as graphs rather than feature vectors.
- Built with
- Python · DeepChem · scikit-learn · NumPy · pandas
Cell-image classifier for malaria, deployed as a mobile-friendly web app.
- Built with
- Python · Keras · scikit-learn · OpenCV · Streamlit
This website — a static TypeScript and Next.js site, built and designed from scratch.
- Built with
- TypeScript · Next.js · React · Tailwind CSS · GSAP
Where I've worked, and what I owned there.
- Ran the build-versus-buy evaluation for the platform and wrote the analysis the decision was made on. We built.
- Trade logging with domain-specific fields, so a result can be interrogated after the fact rather than taken on trust.
- A signal validator that checks strategy output against its own stated assumptions.
- A regime classifier, so performance reads per market condition instead of as one blended number.
- A walk-forward runner — a strategy only tested on the window it was fitted to has not been tested.
Why it is built this way
A backtesting platform makes it cheap to try configurations. That is the risk, not the feature — enough variants selected greedily will produce a clean equity curve on noise alone. The validation layer exists to make that failure mode visible rather than to make results look good.
Stack- Python
- Pine Script
- FastAPI
- PostgreSQL
- Next.js
- Time-series
- Architected and deployed a Python backend implementing the Model Context Protocol with FastMCP, exposing tool APIs for integration into client development workflows.
- Developed a natural-language request layer that converts conversational input into structured tool calls and returns context-aware code output.
Stack- Python
- FastMCP
- Model Context Protocol
- Built reusable, WCAG-compliant React and Next.js UI components adopted across the application's event-booking workflows.
Stack- React
- Next.js
- WCAG
- Designed and deployed containerized Linux infrastructure using Docker and Nginx, reducing deployment time by 60% and integrating transactional email automation.
- Configured SSL/TLS certificates across production web properties to standardize encrypted HTTPS communication.
Stack- Docker
- Nginx
- Linux
- SSL/TLS
Toolkit
Grouped by what I use them for. The numbers point to the projects where each one did real work.