Available for selected buildsportfolio / ai products / full-stack / design systems
I build AI products you can inspect.
Voice, speech and retrieval systems that run on real devices — plus the full-stack surfaces and design systems that make them usable. Nothing hides behind a black box: measured latency, readable state, and an honest list of what is not finished.
Field note
This is not a static resume surface. The site is arranged as an inspectable operating layer: identity, capability, work evidence, and contact routing.
Not a pile of logos. Each row is a working layer of the system — open one to see the stack behind it and where it is proven.
Transcription, synthesis and generation that run on the handset or the laptop — quantised models, VAD gating, and a local model gateway instead of a per-call API bill.
Stack
whisper.cppsherpa-onnxPiperOllamaKokoro
Proven in
Memora voice pipeline
iTantra offline STT/TTS
Chunks local voice I/O
Selected work
Case files.
Problem, thinking, stack, result — the full dossier on one sheet. Most of this is private source; ask and I will walk you through the repo. The terminal stays as an exploration layer for the curious.
On-device AI voice assistant
Memora
A mobile voice-notes app that records meetings and classes, transcribes them with Whisper, and turns the transcript into notes, titles, reminders and exports — in English, Telugu, or both mixed together.
On-device STTCode-switch handlingParallel AI tasks
Problem
Recording a lecture or a client call is easy; getting usable text out of it is not. Cloud transcription breaks on Indian-English-Telugu code-switching, bills per minute, and quietly uploads audio people would rather keep local.
Role
Designed the product, built the Expo/React Native app and the FastAPI pipeline, and wrote the local LLM gateway and schema.
Approach
01Route language to the right model: whisper.cpp base and medium for English, large-v3 for Telugu, selected from the recording's language mode.
02Clean and translate the transcript through a local Ollama gateway — Gemma for generation, Qwen3-Embedding for retrieval — so no per-call API cost is incurred.
03Fan the cleaned transcript out in parallel: title, three note formats, event and deadline extraction, topic tags, and reference links pulled from web search in study mode.
04Keep state honest. Every recording carries an explicit pipeline status in SQLite, and a reminder is created from the exact quoted span of transcript that triggered it.
Outcome
Runs as a daily tool: 36 real recordings processed through the pipeline, exports to PDF and Word, and share-out to WhatsApp or email. Telugu slang accuracy and model cold-start time are written down as open issues, not hidden.
