Ronak Singh

CS student building systems across embedded hardware, applied ML, and infrastructure.

Experience

Software Engineering Intern, Stimulus Consulting Group

Jul 2025 to Aug 2025

Built a customer-facing support chatbot for the company website using JavaScript and the OpenAI API. Implemented few-shot prompting from company documentation and a hybrid routing engine with a 12-turn conversation memory window plus local keyword-fallback logic. Shipped to production in a 14-day sprint. Per client-reported site analytics during the pilot, the bot improved response relevance by about 25% and deflected about 30% of tier-1 support queries.

javascript · openai api

Source

Research

Independent ML Researcher, Assistive BCI Research

Jul 2024 to Nov 2025

Designed and profiled a PyTorch pipeline for P300 brain-computer-interface signal classification, pretraining on synthetic EEG and fine-tuning on real recordings from the MNE dataset. Authored the ethics and limitations analysis covering generalization risk of synthetic-data augmentation in BCI classification. Improved classification accuracy by 6% and cut inference latency 9x, to about 4ms per sample. Published in the National High School Journal of Science. USPTO provisional patent 63/847,808 filed.

pytorch · mne · eeg · bci

Source · Paper

Projects

Open-Source Contributor, hnswlib

Custom HNSW vector database project

Built a C++ HNSW implementation for constrained edge hardware, then backtested it against PostgreSQL and against hnswlib itself, finding hnswlib slower than expected. Traced it to an unfixed open issue in hnswlib's repository, submitted a fix replacing an unreserved push_back loop with a single-allocation range constructor, and measured a 51 to 59% improvement. Verified across CI with ASan and UBSan and merged into develop, credited via commit co-authorship, in a project with more than 5,300 stars used in ChromaDB and Milvus. The own-built database uses int8 quantization, NEON SIMD, and a cache-friendly storage layout, hitting 91.5 to 93.5% recall at 10 and 3.0 to 3.8ms latency on real Raspberry Pi 5 hardware after fixing a graph RNG bug that was causing unstable recall.

c++ · hnsw · raspberry pi · simd · postgresql

GitHub · Merged PR #679

H.A.L.O.: Facial Recognition Access and Alert System

H.A.L.O. demo video

A two-tier facial-recognition access and alert system deployed for real-world household use. A Raspberry Pi (Linux) edge device performs on-device face recognition and writes detection events to Firestore. A Flutter app subscribes in real time through a Firestore listener, and Twilio sends SMS alerts to the homeowner. Diagnosed a reliability bug causing duplicate alerts and engineered an idempotent fix, filed as USPTO provisional patent 63/847,782. Also fixed a data-loss bug where visitor logs were held in an in-memory list that reset on every backend restart, migrating storage to Firestore.

raspberry pi · linux · flutter · firestore · twilio

GitHub · Demo

NFL Elo Ratings and Monte Carlo Outcome Prediction

NFL Elo ratings and Monte Carlo model

Independently validated a 365,541-play prediction pipeline by loading it into PostgreSQL and cross-checking aggregated results against game-level ground truth across 2,743 games. Used SQL joins, GROUP BY aggregations, and window functions to trace a labeling bug: an implausible 8.1% tie rate against the roughly 0.4% actual NFL rate, caused by win and loss labels derived from pre-snap score inference missing the final scoring play. Rebuilt the labels from true final scores, confirming model accuracy at 63.9%, 182 of 285 games. Visualized play-calling trends and model calibration with matplotlib and seaborn, including violin plots, a confusion matrix, and a calibration chart.

python · postgresql · sql · matplotlib · seaborn

GitHub

Autonomous FPV Drone

Autonomous FPV drone

Built and piloted a competition FPV drone, a 4-motor quadcopter with a custom frame and claw-gripper payload, for the TSA Drone Challenge, placing 6th nationally out of more than 140 teams. Configured Betaflight firmware, including rate curves, OSD, and motor-order calibration, and performed iterative PID tuning across three test-flight rounds to eliminate oscillation and drift. After the competition, configured INAV firmware and a GPS module for autonomous follow-me tracking, flight-tested outdoors.

betaflight · inav · gps

GitHub

Gurudwara Library

Gurudwara Library catalog

Volunteer project built with 3 other members: took donated books and built a public library, currently holding more than 200 unique books and more than 330 total books. Handled the technical side, cataloging each book and setting up the cataloging software.

libib

Catalog

Education

Awards

Leadership and Volunteering

Languages

English · Hindi · Punjabi