Skip to content
Arpit Singhal

AI Product & Automation Engineer

I build AI products and automations that work beyond the prototype.

I design and ship agentic workflows, RAG systems, computer-vision applications and production AI backends — from architecture and rapid prototyping through deployment and iteration.

Book a discovery callTell me about your project →

  • i.

    Two live AI mobile apps on the App Store

  • ii.

    Agentic and RAG systems running in production

  • iii.

    AI, backend, deployment and infrastructure ownership

  • iv.

    Generative AI, computer vision, data systems and cloud

Selected work — 01

Free with subscription · Shipped January 2026

Formax workout report for a barbell bench press, showing scores of 8/10 form, 8/10 control, 9/10 range and 7/10 intensity, a colour-coded video timeline, and written tips for improvement.
Fig. 1 — Scored feedback pinned to the moment it happened
Formax app icon

Formax

AI form analysis for lifting videos · v1.4.7

Record or upload a lifting video and Formax scores the movement on form, control, range of motion and intensity — then pins its feedback to the exact moments it happened, on an interactive timeline you can scrub frame by frame.

A subscription product I took from idea to the App Store alone: the vision pipeline, video processing and storage, the mobile experience, the backend, billing, privacy review, and every release since.

  • Video analysis that returns per-segment scores rather than one opaque verdict
  • Timeline feedback tied to specific timestamps, colour-coded by severity
  • AI workout plans generated around training days, equipment and experience
  • Video history and progress reporting across sessions
  • Subscription billing, account deletion and privacy controls

Built with

Expo · React Native · TypeScript · FastAPI · Python · async processing · Google Gemini (video) · Supabase · Google Cloud Storage · GCP · RevenueCat subscriptions

View on App StoreRead the case study →

Selected work — 02

Free with credit-based subscription · Shipped March 2026

COMAIC App Store screenshot showing a photo of a person transformed into illustrated comic panels of a fantasy story.
Fig. 2 — A photo becomes a character, then a multi-page story
COMAIC app icon

COMAIC

Photos and a story idea become illustrated comics · v1.1.0

Add up to five character photos, pick an art style, describe a plot — and COMAIC generates a multi-page comic with panels, dialogue and illustrated pages, in Manga, Manhwa, Comic, Seinen, Shonen, Shojo, Josei and Manhua styles. Up to ten pages per comic, exportable as PDF.

An end-to-end generative-media product: the prompt and image pipeline, asynchronous generation, character consistency across panels, the credit system, PDF export, storage and the App Store release.

  • Photo-to-character generation that keeps a face recognisable across panels
  • Story, dialogue and panel layout generated together rather than separately
  • Eight art styles, each with its own visual treatment
  • Up to ten pages per comic with varied panel layouts
  • Creation history, PDF export and a credit-based subscription

Built with

Expo · React Native · TypeScript · Node · Google Cloud Functions · Vertex AI (image generation) · Firebase · Cloud Storage · RevenueCat, credit metering

View on App StoreRead the case study →

Services — what I take on

How I can help

Three ways clients usually bring me in. Each one is scoped as an outcome with a working system at the end of it — not a research exercise.

01

AI Agents & Workflow Automation

Agents that actually touch your systems — reading, deciding and acting across the tools your team already uses, with a human in the loop wherever a mistake would be expensive.

  • Tool-using agents wired into real APIs
  • Multi-agent and stateful workflows
  • Email, CRM, database, scheduling and API integrations
  • Human-in-the-loop review and approval gates
  • Monitoring, retries and reliability work

Right fit if

  • A workflow eats hours of skilled time every week
  • The steps are well understood but tedious and error-prone
  • You need an audit trail, not a black box

02

RAG & Knowledge Systems

Assistants that answer from your documents and data instead of guessing — built with the retrieval evaluated, because a confident wrong answer is worse than no answer.

  • Document ingestion and retrieval pipelines
  • Internal knowledge and customer-support assistants
  • Retrieval evaluation and hallucination reduction
  • Access control, so people only retrieve what they may see
  • Admin tooling and usage analytics

Right fit if

  • Answers live in documents nobody can find
  • Support repeatedly answers the same questions
  • You tried a generic chatbot and it made things up

03

End-to-End AI Product Engineering

A prototype that works on your laptop is not a product. I take AI features the rest of the way — backend, infrastructure, billing, release — and keep shipping after launch.

  • Generative AI and computer-vision features
  • FastAPI and production Python backends
  • Async processing and task queues
  • Mobile and web product integration
  • Docker, cloud deployment, CI/CD and monitoring

Right fit if

  • A demo impressed people and now has to become real
  • You need one person who owns AI, backend and deployment
  • You are shipping to real users and cannot hand-wave reliability

Client work — anonymised

Selected client work

Work delivered for employers and clients, described by sector and outcome. Implementation details stay with the people who own them.

Client and employer names are withheld. Figures below are the measurements recorded on those projects.

01 · Operations intelligence

Time to investigate

daysunder 30 minutes

Investigating business anomalies without a human trawling the data

Challenge
When a location or metric moved unexpectedly, finding out why meant an analyst manually cross-referencing internal records against whatever was happening in the outside world. It took days, and it happened often enough to occupy real headcount.
What I built
An AI agent that correlates patterns in the internal database against local news and external event signals, forms candidate explanations for an anomaly, and presents them with the evidence it used.
Engineering considerations
  • The agent proposes explanations; a person still decides — it is an investigator, not an authority
  • Every conclusion carries the sources it came from, so it can be checked
  • External feeds are unreliable and the workflow has to degrade rather than fail

Python · LLM agent · SQL · external news & event APIs

02 · Geospatial & location data

Model accuracy

29%86%

Lifting model accuracy by fixing the inputs, not the model

Challenge
A prediction model was performing badly enough to be unusable. The limiting factor was not the algorithm — it was that the signal it needed was spread across several unrelated external sources.
What I built
A pipeline that aggregates and reconciles signals from maps, weather, traffic and other external APIs into a single feature set, handling disagreement and gaps between sources.
Engineering considerations
  • Third-party APIs fail, rate-limit and contradict each other; ingestion assumes it
  • Features are reproducible, so a score can be explained after the fact
  • Most of the accuracy came from data engineering, not model selection

Python · ML pipeline · maps / weather / traffic APIs

03 · Internal data tooling

Internal analytics throughput

baseline3× faster

Letting the team query the database in plain language

Challenge
Routine analytics and data changes queued behind whoever could write the SQL, which made small questions expensive and slow.
What I built
An LLM-powered MCP agent that performs PostgreSQL create, read, update and delete operations from natural language, scoped to what the requesting user is allowed to touch.
Engineering considerations
  • Write operations are the dangerous ones and are gated accordingly
  • Generated SQL is inspectable before it runs
  • Permissions follow the user, not the agent

Python · MCP · PostgreSQL · LLM tool calling

04 · Commerce & customer communication

Customer journey

four separate handoffsone workflow

One agent graph across discovery, policy, booking and email

Challenge
Customer questions spanned product discovery, company policy, scheduling and follow-up email — four different jobs that customers did not experience as separate.
What I built
A LangGraph multi-agent workflow covering product discovery, company policies, meeting booking and automated email replies, plus a voice-calling agent built on the OpenAI Realtime API and Twilio.
Engineering considerations
  • Shared state across agents, so the customer never repeats themselves
  • Voice adds latency and interruption handling that text never has to solve
  • Anything that sends on the company’s behalf needs a confidence floor

LangGraph · OpenAI Realtime API · Twilio · Python

Engagement model

How engagements run

Most projects start small on purpose. You should be able to find out whether this works — and whether we work well together — before committing to anything large.

  1. 1

    Discover

    Identify the workflow, the business outcome, the constraints, the data that actually exists, and the risks. Often this is where a project gets smaller and sharper.

    • Problem definition
    • Data and systems inventory
    • Risk and constraint list
  2. 2

    Validate

    Define a bounded pilot with measurable acceptance criteria, so success is something we agreed on in advance rather than argued about afterwards.

    • Scoped pilot
    • Acceptance criteria
    • Fixed timeline and cost
  3. 3

    Build & deploy

    Implement the system with observability, security and operational reliability built in — then actually put it in production, not in a demo branch.

    • Working system in production
    • Monitoring and alerting
    • Handover docs
  4. 4

    Improve

    Evaluate real usage, tighten quality where it matters, and extend the workflows that are earning their keep.

    • Usage evaluation
    • Quality improvements
    • Extension roadmap

We can start with a focused technical discovery or a paid pilot rather than a large open-ended commitment.

About

About

I am an AI engineer working across generative AI, agentic workflows, RAG, computer vision, data engineering, backend systems and cloud deployment.

What I actually do is own the whole lifecycle. I prototype, decide the architecture, build it, deploy it, and keep improving it once real people are using it. That end-to-end ownership is the reason my own apps are on the App Store rather than in a folder of experiments.

It also changes how I work with clients. I have shipped, priced, supported and iterated products of my own, so I tend to ask about cost, failure modes and what happens in month three — not just whether the model output looks impressive in a demo.

GitHubLinkedInApp Store developer page

Capabilities

Languages
Python · JavaScript · TypeScript · SQL
AI & agents
LangGraph · LangChain · OpenAI · Gemini · Hugging Face
Backend
FastAPI · Django · Celery · Redis
Data
PostgreSQL · Supabase · Firebase
Infrastructure
Docker · AWS · GCP · Azure · Nginx · Cloudflare

Contact

Tell me what you are trying to build or automate.

Send a few lines about the workflow or product you have in mind. If it is a fit, I will come back with how I would approach it and what a first phase would look like. If it is not, I will say so.

Or book a call

Book a discovery call

Scheduling link coming soon — the form is the fastest way to reach me in the meantime.

Optional — how it works today, and what it costs you

I reply to every genuine inquiry, usually within two business days.

Prefer email? Write to arpit@arpitsinghal.com.