Interested in building systems from scratch
-
ProductLabs AI Form Filler
Real-time information extraction pipeline to fill structured forms from live audio/text conversations.- Built a multimodal media layer running Whisper ASR and PyAnnote acoustic embeddings to automate multi-speaker diarization with under 200 ms chunk ingestion latency.
- Implemented a custom worker thread pool to balance data translation tasks and Qwen-2.5-1.5B inference, dropping peak response latency by 65%.
-
FastAPIPythonMongoDBQwen-2.5PlaywrightPyAnnoteWhisper ASR
-
AI-Powered Tutor
Offline fully personalized AI-based tutor running on the Qualcomm Innovators Dev Kit.- Optimized model inference routines directly on the edge hardware NPU via QAIRT.
- Pushed compilation throughput from
$\sim 3$ tokens/s to 26–30 tokens/s and slashed Time-To-First-Token (TTFT) down to 25 ms. -
Android (Java)QAIRTGenie SDKLLaMA-3.2-3BQualcomm NPU
- Cricket Match Timeline Generator
CV and OCR pipeline that parses raw cricket videos into ball-by-ball text timelines.- Trained two custom ResNet18 models (modified for 1-channel grayscale) to track bowler action and predict exact ball-release timestamps.
- Paired PaddleOCR with a custom state machine to read the broadcast scoreboard at those specific frames and infer missing data like wides or occluded scores.
- Built an interactive CLI tool with smart caching to avoid running heavy GPU inferences twice.
PyTorchOpenCVPaddleOCRPython
-
Distributed File System
High-reliability network file system supporting concurrent client read/write access.- Replaced classic linear file path scans with a thread-safe Trie-based index, optimization complexity down to
$O(L)$ for a 3–4$\times$ resolution speedup. - Enforced operational stability using sentence-level locks, fault-tolerant replication checkpoints, and automated tracking maps.
-
CTCP/IP StackPOSIX Threads
- Replaced classic linear file path scans with a thread-safe Trie-based index, optimization complexity down to
-
CLI-Shark
Terminal-based high-throughput packet sniffing tool for real-time network.- Integrated BPF filtering directly at the kernel layer to drop noisy frame processing early, reducing CPU load metrics by
$\sim$ 60 %. - Parses Ethernet, IP, and TCP/UDP transport wrappers on raw operational boundaries.
-
ClibpcapPOSIX NetworkingTCP/IP
- Integrated BPF filtering directly at the kernel layer to drop noisy frame processing early, reducing CPU load metrics by
- AmbientSynth
Cross-platform terminal-based MIDI piano synthesizer with real-time audio effects.- Dynamic audio backend selection for seamless playback across macOS, Linux, and Windows.
- Supports live adjustment of sustain, reverb, chorus, and expression parameters during performance.
- Built with FluidSynth for SoundFont synthesis and mido for low-latency MIDI input handling.
PythonFluidSynthSoundFontMIDI
