yash bhalgat

a childhood portrait of yash bhalgat in a warm orange and yellow collage

founding engineer @ stealth startup

i build thoughtful interfaces
and the systems behind them.

give your ai excellent taste in engineers.

github linkedin email

los angeles, california

where i’ve worked.

experience
stealth startupfounding engineer · dec 2025 — presentai-assisted film post-production, captioning, and secure video delivery.

i build the systems that take a film from raw footage to broadcast-ready delivery, covering audio understanding, caption quality, streaming, and rendering.

captioning and quality control

a multi-pass gpu pipeline combines vocal isolation, whisperx transcription, forced alignment, speaker diarization, sound-event detection, and llm authoring with an automated judge. a deterministic quality-control engine checks reading rate, line length, and broadcast timing across 12 caption formats, with every output verified against a real decoder.

secure playback and packaging

an authenticated streaming gateway, widevine and fairplay drm, short-lived sessions, per-viewer forensic watermarks, and access audit records protect pre-release footage. parallel, keyframe-aligned transcoding supports feature-length packaging, adaptive streaming, and seeking.

delivery tools

an end-credits builder renders through a headless browser and ffmpeg to broadcast-spec mp4, prores, and dpx. i also built a film tax-incentive compliance panel and engineering controls for soc 2 readiness.

  • python
  • whisperx
  • ffmpeg
  • llm orchestration
  • secure streaming
  • soc 2 readiness
smartpractice aifounding engineer · jun — dec 2025clinical documentation, medical coding, and ehr integration in one app.

an ambient clinical documentation app that turns encounter audio into reviewable notes and coding suggestions. clinicians can use standard templates or define their own format, review the evidence, and send the result into their ehr.

from conversation to clinical note

streaming speech recognition uses medical-lexicon biasing and separates clinician, patient, caregiver, and interpreter roles. each generated assertion cites its transcript span. context classification distinguishes negation, hypothetical statements, and family history from active patient problems.

custom note formats

clinicians compose a format section by section or upload an existing note as an example. the format compiles into a typed extraction schema with grounding rules and loinc document types. one engine supports soap, apso, h&p, procedure notes, dap/birp behavioral health notes, and pt/ot evaluations.

evidence-backed coding

icd-10-cm, cpt, and hcpcs suggestions include supporting quotes. the coding engine covers e/m medical-decision-making levels, modifiers, ncci bundling, diagnosis-to-procedure linkage, and hcc recapture with meat evidence. accepted suggestions become a superbill.

ehr interoperability

smart on fhir supports ehr launch, standalone launch, and backend services. notes are written as documentreference resources alongside structured clinical data, with snomed ct, icd-10-cm, and rxnorm terminology mapping. fhircast keeps the app aligned with the clinician’s open chart.

working across ehr systems

a manifest v3 chrome extension provides a write layer where vendors do not expose write apis. it uses per-system permissions, short-lived scoped tokens, no phi stored in the extension, and a per-field trail linking each entry to its clinician, note, and timestamp.

access and accountability

role-based access, mfa, recording-consent capture, segmented substance-use content, and append-only logs cover sensitive clinical workflows. i owned the engineering controls supporting the company’s soc 2 type i and type ii audits.

  • react
  • typescript
  • supabase
  • postgresql
  • llms
  • smart on fhir
  • hipaa-compliant workflows
  • soc 2 type i & ii

things i’ve built.

select to explore
inkcastai video creationa topic becomes a narrated, hand-drawn lesson — with every frame compiled from vectors.

turn a question, document, or web page into a whiteboard explainer with diagrams, narration, and captions. a readable storyboard is the source of truth for every frame.

  • speech-synced diagrams
  • deterministic rendering
  • local voice models

drawings that follow the explanation

a seven-stage compiler plans the lesson, creates a structured storyboard, measures kokoro narration, and aligns words with whisper. diagrams appear when they are mentioned. chromium renders vector strokes, equations, and plots; ffmpeg assembles the final video and captions.

creative control, with an offline path

choose a format, drawing style, and narration, then preview, download, or share the finished film. fireworks powers optional lesson authoring, critique, vector diagrams, and visual review. local planning and speech also run without provider keys; lesson and storyboard editing are available in the cli and self-hosted studio.

a separate rendering service

the next.js studio handles identity and dispatch; a stateless vercel container renders one film per request. a shared private blob store holds output and progress, streamed to the browser over sse. scene-level caching and parallel render segments keep local re-renders efficient, while authenticated media routes enforce access.

how it fits together
  1. topic / document
  2. lesson + storyboard
  3. voice + quality checks
  4. 1080p video
architecture & diagrams on github private

current scopehosted generation, download, and sharing; editing and re-rendering are currently self-hosted features.

  • next.js
  • typescript
  • fireworks
  • kokoro / onnx
  • whisper
  • chromium
  • ffmpeg
  • vercel containers
  • private blob
  • sse
inkcast project details inkcast.video github private
tallywebaccounting & erp · tally-updatedkeyboard-first books, inventory, and reports with accounting rules enforced in the database.

a browser workspace for indian businesses and accountants: enter vouchers, track stock, reconcile bank statements, and drill into financial reports using familiar keyboard navigation.

  • double-entry ledger
  • versioned vouchers
  • company-scoped access

the books, from entry to report

sales, purchases, receipts, payments, and ledgers feed the day book, balance sheet, profit and loss, and outstanding balances. inventory covers stock movements, orders, and bills of materials. bank imports, reconciliation, gst reports, and tds workflows share the same workspace.

correctness below the interface

the go core is the only write path. postgres enforces double entry, versioned and hash-chained vouchers, and default-deny force rls. a per-company policy engine checks people and delegated agents before a transaction. an offline outbox queues entries; advisory locks and unique run keys protect background work.

ai with an approval boundary

the architecture supports natural-language queries, an approval inbox, mcp clients with scoped grants, and six model-provider adapters. a separate python service turns bills into purchase-voucher drafts after ocr and pii redaction. proposals must pass the same permissions and accounting rules as human actions.

deployed on aws

vercel serves the react app; cloudfront routes the api to an arm64 go lambda through the lambda web adapter. aurora serverless v2 stores the books, pgvector supports retrieval, and eventbridge scheduler drives idempotent background jobs. terraform defines the deployment.

how it fits together
  1. react + xstate
  2. go accounting core
  3. policy + transaction
  4. aurora postgres
architecture & diagrams on github private

current scopethe accounting core is deployed on aws. live model providers, the bill-processing service, and statutory filing integrations are not yet configured in that deployment.

  • go
  • react
  • typescript
  • xstate
  • postgresql
  • force rls
  • pgvector
  • fastapi
  • aws lambda
  • cloudfront
  • aurora
  • eventbridge
  • terraform
  • mcp
tallyweb project details tallyweb.world github private
finflowwfinancial analysis · finitoreconcile accounts and investigate uploaded data with an evidence trail back to the source row.

a finance workspace for bank reconciliation, invoice-to-purchase-order review, budget-versus-actual analysis, and conversations over uploaded files. deterministic code computes money; models help match, investigate, and explain.

  • source-row citations
  • bounded code execution
  • human approval gates

three recurring finance workflows

compare erp exports with bank statements, match invoices to purchase orders, and prepare budget-versus-actual commentary. a dependency-graph orchestrator runs ready steps concurrently, records their inputs and model provenance, retries failures, and pauses payable-invoice decisions for a reviewer.

investigations you can inspect

uploaded spreadsheets, json, text, and embedded-text pdfs become immutable, hashed sources with original row numbers. a bounded agent explores them through a closed tool registry and can write calculations in an isolated quickjs sandbox with no filesystem or network access. final figures are recomputed in typescript from cited rows.

history, memory, and review

append-only events capture each investigation turn and tool result, with realtime updates to the browser. postgres full-text search retrieves prior sources and workflow history. role-based approval rules, ceilings, and optional separate approvers keep consequential proposals under human control.

one deployable, tenant-aware data

next.js on vercel connects to supabase auth, postgres, private storage, and realtime. server handlers enforce organization access; rls scopes browser reads. organizations can bring their own model providers, with api keys encrypted using aes-256-gcm.

how it fits together
  1. uploaded sources
  2. workflow / investigator
  3. evidence + approvals
  4. reviewable report
architecture & diagrams on github private

current scopepilot software. background runs currently use function lifetimes; a durable queue and automatic recovery remain production-readiness work.

  • next.js
  • react
  • typescript
  • supabase
  • postgresql
  • realtime
  • quickjs / wasm
  • zod
  • agent orchestration
  • fireworks
  • github actions
finfloww project details finfloww.xyz github private
dentabilldental claims & voice aia claim-follow-up workspace with live transcripts, structured outcomes, and a next-action plan.

dental billing teams import unpaid claims, queue follow-ups, and review call transcripts and structured summaries. the current demo exercises the real application pipeline with simulated payer calls.

  • live transcript stream
  • queued claim workflows
  • audited phi access

from an unpaid claim to a next step

csv imports create organization-scoped claims. the dashboard follows queued calls, streams transcript events, and presents a structured claim outcome and recommended action. summaries use versioned prompts and validated schemas, with a transcript-derived fallback in the demo.

reliable background work

a fastify api sends call and summary jobs to sqs. workers use conditional claim-state updates and unique summary records to reduce duplicate effects. retries back off, failed calls move to manual review, and a dead-letter queue captures unhandled failures. the same handlers support bullmq locally.

streaming across serverless infrastructure

the next.js frontend runs on vercel; cloudfront and the lambda web adapter serve the api in response-streaming mode. transcripts replay from postgres after reconnects. separate worker lambdas, aurora serverless v2, private subnets, cloudwatch, and terraform complete the aws deployment.

organization boundaries and audit

clerk jwt verification establishes the organization. prisma write guards, scoped reads, and phi access audits enforce tenancy, while signed webhooks and redacted logs protect integrations. live retell telephony remains gated on a baa, ivr tooling, and completed duration and hold controls.

how it fits together
  1. claim import
  2. sqs + call worker
  3. summary + state transition
  4. transcript / next action
architecture & diagrams on github private

current scopelive demo in simulation mode: no insurance payer is actually dialled. the queue, worker, database, transcript, claim transitions, and audit path run end to end.

  • next.js
  • typescript
  • fastify
  • prisma
  • clerk
  • aws lambda
  • sqs / dlq
  • aurora
  • sse
  • terraform
  • voice agents
  • claude
dentabill project details usedentabill.xyz github private
cursor for resolvevideo editingon-device ai assistance inside a professional edit bay.

an agent that works with davinci resolve for dailies, screener exports, and vfx pulls, alongside video matting and promptable segmentation.

local by design

an on-device vision-language model keeps source footage on the machine. the agent connects visual understanding with editing actions inside resolve.

  • computer vision
  • on-device ai
  • davinci resolve

what i work with.

skills & systems
aiagents, speech, and grounded generation

lesson planning and visual review in inkcast; bounded investigations in finfloww; delegated agents in tallyweb; voice workflows in dentabill. medical speech, transcript grounding, and multi-pass evaluation in my founding-engineer work.

  • llm orchestration
  • tool calling
  • rag
  • mcp
  • structured outputs
  • evaluation harnesses
  • fireworks
  • claude
  • openai
  • gemini
  • kokoro / onnx
  • whisperx
  • diarization
  • forced alignment
  • vector-video compilation
cloudaws, gcp, and managed deployments

serverless apis and workers, private data services, scheduled jobs, and infrastructure as code. inkcast separates its studio and renderer; tallyweb and dentabill share aws infrastructure with separate databases and roles.

  • aws lambda
  • cloudfront
  • aurora serverless v2
  • sqs
  • eventbridge
  • s3
  • ec2
  • cloudwatch
  • secrets manager
  • iam
  • private vpcs
  • gcp cloud run
  • gcs
  • vertex ai
  • vercel containers
  • supabase
  • docker
  • kubernetes
  • terraform
  • github actions
backenddata models, apis, and business rules

transactional accounting rules, organization-scoped data, typed apis, webhook verification, and event-driven workflows. database constraints enforce invariants beneath the application layer.

  • go
  • python
  • typescript
  • node.js
  • fastify
  • express
  • django
  • fastapi
  • postgresql
  • prisma
  • redis
  • mongodb
  • rest / openapi
  • websockets
  • sse
  • sql transactions
  • rls
  • pgvector
  • zod
frontendresponsive interfaces and live workflows

keyboard-first accounting, live claim transcripts, investigation timelines, and creative studios. explicit state machines and incremental updates keep complex workflows understandable.

  • react
  • next.js
  • react server components
  • typescript
  • javascript
  • tailwind css
  • xstate
  • redux
  • vite
  • base ui
  • shadcn/ui
  • accessibility
  • keyboard navigation
scalabilityconcurrency, queues, and failure recovery

scale-to-zero deployments, isolated render jobs, cached scene segments, concurrent workflow steps, and replayable event streams. i make retry behavior, resource limits, and recovery paths explicit.

  • distributed systems
  • idempotency
  • sqs / dlq
  • bullmq
  • worker leases
  • retry backoff
  • bounded concurrency
  • pub/sub
  • scene caching
  • parallel transcoding
  • database indexing
  • connection budgets
  • vitest
  • playwright
soc 2 complianceengineering controls and audit evidence

owned engineering controls supporting smartpractice ai’s soc 2 type i and type ii audits. at stealth startup, built security and confidentiality controls for type i readiness, including tenant rls, audit trails, and security checks in ci.

  • tenant isolation
  • least privilege
  • append-only audit logs
  • mfa
  • ci security scans
  • access reviews
  • risk reviews
  • encryption
  • secret management
hipaa-compliant workflowsclinical data, access, and interoperability

at smartpractice ai, built clinical workflows with role-scoped access, mfa, recording consent, phi audit records, and no phi stored in the ehr browser extension. clinical notes and coding suggestions retain transcript evidence and clinician review.

  • phi access auditing
  • minimum-necessary access
  • role-based access
  • consent capture
  • 42 cfr part 2 segmentation
  • hl7 fhir r4
  • smart on fhir
  • fhircast
  • loinc
  • snomed ct
  • rxnorm
  • icd-10-cm
  • cpt / hcpcs

a little background.

educationcomputer science at uc riverside and pict.

university of california, riverside

m.s. computer science · 2023 — 2025
gpa 3.76 / 4

graduate teaching assistant for physics and chemistry labs, with 45 students per term.

pune institute of computer technology

b.e. computer engineering · 2019 — 2023
gpa 9.2 / 10

let’s make something
worth using.

download résumé

on repeat lately.

open spotify
let’s talk