I build thoughtful interfaces and the systems behind them: creative tools, AI pipelines, and dependable backends.
Founding Engineer, Stealth Startup · Dec 2025 – present AI-assisted film post-production: captioning, secure video delivery, and rendering.
- Multi-pass GPU captioning pipeline (vocal isolation, WhisperX, forced alignment, diarization, sound-event detection, LLM authoring with an automated judge) that cut feature-film captioning from a multi-day vendor cycle to one automated run.
- Deterministic caption QC engine with auto-fixers over 12 delivery formats (SCC, TTML/IMSC, iTT, EBU-TT-D, SRT, WebVTT), each file verified against a real decoder.
- Secure pre-release playback: authenticated streaming gateway, Widevine/FairPlay CMAF, per-viewer forensic watermarking, and a parallel keyframe-aligned transcode for adaptive streaming.
- End-credits render pipeline to broadcast-spec MP4, ProRes and DPX; led SOC 2 readiness to Type I.
Founding Engineer, SmartPractice AI · Jun – Dec 2025 Ambient clinical documentation, medical coding, and EHR integration.
- Encounter audio → streaming ASR with medical-lexicon biasing → speaker roles → LLM-structured notes where every assertion cites its transcript span.
- Note-format engine that compiles a clinician's own format into a typed extraction schema (SOAP, APSO, H&P, DAP/BIRP, PT/OT on one engine).
- Evidence-backed ICD-10-CM / CPT / HCPCS coding with E/M levels, NCCI bundling and HCC recapture, exported as a superbill.
- SMART on FHIR integration (EHR, standalone and backend-services launch, FHIRcast) plus an MV3 Chrome extension as a write layer where EHRs expose no write API.
- Engineering controls for HIPAA and SOC 2 Type I and Type II.
| Project | What it is | Stack |
|---|---|---|
| Inkcast | A topic or document becomes a narrated, hand-drawn vector lesson. A seven-stage compiler syncs drawings to measured speech and renders with Chromium and FFmpeg in stateless containers. | TypeScript, Next.js, Kokoro/ONNX, Whisper, Fireworks |
| Tallyweb | Keyboard-first accounting and ERP with double-entry constraints, hash-chained vouchers and FORCE RLS in Postgres. Go core on ARM64 Lambda behind CloudFront, Aurora Serverless v2. | Go, React, XState, Terraform, AWS |
| Finfloww | Reconciliation, invoice-to-PO review and variance analysis with deterministic calculations, human approval gates, and sandboxed (QuickJS) investigations that cite source rows. | Next.js, TypeScript, Supabase |
| DentaBill AI | Dental claim follow-ups: imports, queued payer calls, live transcripts and structured next actions. Fastify API and SQS workers on Lambda with Aurora and PHI audit trails. | Next.js, Fastify, Prisma, Clerk, AWS |
| Cursor for Resolve | An agent inside the edit bay: video matting, promptable segmentation and an on-device vision-language model driving DaVinci Resolve. No footage leaves the machine. | Python, on-device VLMs |
More detail, demos and architecture write-ups on yashbhalgat.com.
Languages: TypeScript, Python, Go, SQL, C++ Backend & data: Node.js, Fastify, FastAPI, Django, PostgreSQL (RLS), Redis, MongoDB, Supabase, event-driven pipelines Frontend: React, Next.js, XState, Redux, Tailwind CSS Cloud & DevOps: AWS (Lambda, SQS, Aurora, CloudFront, EventBridge), GCP (Cloud Run, Vertex AI), Docker, Kubernetes, Terraform, GitHub Actions AI: LLM orchestration (Claude, Gemini, OpenAI), agentic tool use, RAG, MCP, evaluation harnesses Healthcare: HL7 FHIR R4, SMART on FHIR, ICD-10-CM / CPT, HIPAA, SOC 2
- M.S. Computer Science, University of California, Riverside · 2023 – 2025 · GPA 3.76/4 · Graduate Teaching Assistant
- B.E. Computer Engineering, Pune Institute of Computer Technology · 2019 – 2023 · GPA 9.2/10
- Spatial-keyword queries, ACM SIGSPATIAL
Fun fact: I speak fluent German.