I work on speech and vision systems that run on ordinary hardware.

A voice agent that makes and takes real phone calls for a real-estate client. Audio streams from Twilio through speech-to-text, an LLM, and back out as speech, fast enough to feel like a conversation rather than a form.

• Sub-second round trip, and it stops talking when you interrupt it
• Cut token spend ~40% by compressing conversation state instead of resending it
• Swappable provider layer, so a new model takes hours rather than days

• Deepgram
• Llama 3.3-70B
• Sarvam TTS
• MongoDB
• WebSockets

Assistive communication for someone who can't speak or move much: one webcam watches which registered object they're looking at, and tells a caregiver. No calibration routine, because a patient in a bed can't perform one.

• A detector I trained for this scene runs 330× faster than the zero-shot baseline
• Benchmarked 8 model variants to find one that holds up on CPU alone
• Abstains rather than guessing — a wrong word is worse than no word

• Gaze-LLE
• YOLO-World
• Depth Anything V2
• OpenCV

An email outreach platform. Sequences send from real mailboxes on a schedule; when someone replies, the sequence stops, the reply gets classified, and it becomes a CRM entry.

• Sending exactly once is enforced by database constraints, not by hoping the worker behaves
• A background worker does the sending, so campaigns run with no browser open
• Per-mailbox daily caps and send windows, in the mailbox's own timezone

• Bun
• React 19
• Prisma
• Tailwind
• Playwright

A campus app for my college. Events from every club in one place, and registration that fills itself in from your profile — the specific annoyance of retyping your roll number for the fifth time.

• 85 tests passing, analyzer clean, running on a real device against live Postgres
• Permissions live in row-level security policies, so the database refuses bad reads
• Club admins get a role in the same app, not a second one to install

• Dart
• Riverpod
• go_router
• PostgreSQL

I was losing evenings to checking scholarship and internship listings, so I stopped. This reads twelve sources, throws out anything irrelevant to me, and sends the rest to Telegram.

• About 170 listings a run, deduped against everything already sent
• A small model does the relevance call, which is cheaper than my attention
• Runs on a cron in Actions; no server to keep alive

• Groq
• Llama 3.1
• Telegram API
• RSS

Autism classification datasets are small. A WGAN-GP learns the shape of functional-connectivity matrices and generates more, to see whether synthetic training data genuinely helps.

• The test set is split off before the GAN sees anything — otherwise the result is meaningless
• 200×200 correlation matrices flattened to 19,900 features
• Baseline and augmented classifiers judged on the same untouched holdout

• WGAN-GP
• fMRI
• Jupyter

A POSIX-compliant shell written by hand for the CodeCrafters challenge. Quoting rules are far stranger than I expected before writing the parser.

• REPL, command parsing, PATH resolution, redirection, builtins
• No shell libraries — process handling done directly

• POSIX
• subprocess
• uv

Moving the mouse by looking, using a plain webcam and no neural network. Facial landmarks find the eyes, thresholding finds each iris, and two calibration points map gaze to the screen.

• Contour detection for the iris centre — classical vision, and fast
• Groundwork for the gaze estimation in GAZE

• dlib
• pyautogui
• NumPy

Upload a résumé, get back the skills and seniority it reads from it, plus live job listings matched against them. Nothing is stored anywhere.

• Gemini pulls out skills, experience level and location preference
• Matched against the APIJobs feed with direct apply links

• Streamlit
• Gemini
• PyPDF2