All work

TaxBlitz — US 1099-series and India ITR filing automation. Azure Document Intelligence extraction, nine selectable LLM providers, and human-in-the-loop gates so no AI-derived figure reaches tax math unreviewed.

Cyloid — Shared document-intelligence and AI-governance layer — hash-chained audit trail, Z3 SMT verification as a governance signal, and reversible PII pseudonymisation for DPDPA compliance.

Acc'ademy — AI-native learning platform: a three-role tutor orchestrator over six tenant-scoped tools, a no-code DAG agent builder, and citation-integrity enforcement that strips any evidence marker the model didn't actually retrieve.

AI Dubbing — Video dubbing into 20 languages — Deepgram transcription, GPT-4o translation, and a self-hosted XTTS v2 service that clones the source speaker from a 30-second reference clip where the language supports it.

AI Clipper — Turns long videos into captioned vertical highlights. Word-level timestamps drive burned-in subtitles, and clips are cut with FFmpeg stream copy so there's no re-encode.

AI Ad Generator — Product image to rendered video ad. Vision analysis feeds script generation behind a swappable provider interface, with per-provider circuit breakers, idempotency keys and graceful WebSocket-to-SSE degradation.

Video Explainer — CLI that turns a rough screen recording into a polished multi-language explainer. A locked vocabulary keeps on-screen terms verbatim through rewriting and translation, and video is only ever cut, never slowed.

Meeting Note-Taker — Headless bot that joins calls as a real participant, transcribes per-speaker, and survives crashes by rehydrating from persisted segments. Adaptive map-reduce summarization with a per-chunk failure floor.

Voice Receptionist — Realtime WebRTC voice agent with database-backed appointment booking. A barge-in gate only unlocks interruption after repeated attempts, so echo and room noise don't cut the agent off mid-sentence.

Voxtral Plugin — Streaming STT and TTS provider plugin for telephony, built on the batch endpoint rather than the realtime socket — fully curl-testable, no WebSocket dependency, and about 2x cheaper per minute than Whisper for speech-to-text.