Open to Forward Deployed Engineer roles

I take AI out of the demo and into the hands of the people who use it.

AI engineer with two years of shipping production LLM systems for clients — multi-agent platforms, retrieval over company data, real-time voice. Based in Ahmedabad. Open to remote work and relocation.

See the work

~/career.logtail -f

2022-06BSc, AI & Machine Learning — Gujarat University
2024-04MSc, AI & ML. Invoice extraction (YOLO + OCR) at Chaitanya AI
2024-09Joined Estatic Infotech as AI/ML developer
2025Shipped an AI assistant for financial advisors at a fintech client
2026-08#127 of 1,982 in the HackerRank Orchestrate hackathon
nowLooking for the next team to deploy with

#127 of 1,982: HackerRank Orchestrate, a 24-hour agent-building hackathon
2 years: shipping AI systems for clients
33: projects across agents, data, voice and vision

Selected work

Three projects, each from a different side of the job: competing under a deadline, running a client system in production, and making company data answer questions.

01 · Competition · #127 of 1,982

Message Notification Router

Built in 24 hours for HackerRank Orchestrate. A WhatsApp stream mixes family chats, notices, promotions, posters, voice notes and scams. My system decides for each message whether to notify, digest or mute, says why, and cites past messages as evidence.

Design: A generator agent drafts each decision. A reviewer agent checks it against a fixed five-point list.
Safety: Prompt injection, OTP phishing and lookalike domains are muted, and review can't talk a scam back up.
Stack: Python · Groq Whisper · vision OCR · strict JSON schema · retrieval over history

Leaderboard

message · text, image or voice

context · sender, history, past reactions

generator agent

reviewer agent

rejected drafts loop back, with guards

notify
digest
mute

02 · Client work · Fintech · In production

An AI assistant for financial advisors

Advisors ask in plain language or by voice, in Indian languages. It researches and compares mutual funds, analyses client portfolios, runs SIP, STP and SWP calculations, builds PDF reports, drafts email and WhatsApp campaigns, and surfaces leads.

Role: AI engineer and main client contact: requirements, build, updates
Design: An LLM orchestrator picks from 40+ domain tools. A dispatcher routes each request across model providers by cost and difficulty.
Scale: Multiple workers share sessions and events through Redis pub/sub, with conversations flushed to SQL Server.
Stack: FastAPI · Socket.IO · SQL Server · ChromaDB · Redis · OpenAI · Groq · Sarvam (voice)

Client and product names withheld.

advisor · text or voice

authenticated, encrypted real-time connection

dispatcher · picks the model

orchestrator · 40+ tools

fund research
recommendations
portfolios
calculators
reports
campaigns
leads
insurance

SQL ServerRAG · ChromaRedis sessionsclient APIsTTS / STT

03 · Client work · Enterprise data

Asking company databases questions in plain language

Query engines for an employee-management system on MS SQL and a learning-management system. Staff ask in plain English. The engine picks the relevant tables from the schema, writes the SQL, runs it, and returns a readable answer. It remembers the conversation for follow-up questions.

Role: Design and build, from schema to interface
Stack: LangChain · LlamaIndex · Azure OpenAI · Groq · SQLAlchemy · Flask

illustrative

askWhich departments hired more than five people this quarter?

tablesemployees, departments

SELECT d.name, COUNT(*) AS hires
FROM employees e
JOIN departments d ON d.id = e.dept_id
WHERE e.joined_on >= :quarter_start
GROUP BY d.name
HAVING COUNT(*) > 5;

answer3 departments — Engineering (12), Sales (8), Support (6).

What I'm building for myself

Tools I want to exist, and where each one stands today.

Ummul Kitab
In development

A Quran and Islamic-knowledge study platform: translations, tafsir, word-by-word meaning, IndoPak mushaf pages, a hadith library in Arabic, Urdu and English, and a sourced Seerat timeline.

Next.js 16, FastAPI, PostgreSQL
Haya
In development

An on-device Android content filter. It detects and blurs selected content on screen in real time, with no network calls. I benchmarked the quantization options and chose fp32 at 320×320; on-phone latency is the next thing to measure.

Kotlin, TFLite, YOLO segmentation
15Hisaab
In development

A private 15-minute time log for Android. Every block of the day gets one honest line. Encrypted on the device, with no accounts and no network permission.

Flutter, Drift, SQLCipher
Trivium
Early stage

A side project exploring how developers can keep what they learn while coding with AI. Still an experiment.

FastAPI, React, TypeScript

How I work with clients

Find the real problem
The first request is rarely the real need. I ask until I can explain the user's day back to them.
Prototype on their data
A rough version on real data within days beats a polished one built on assumptions.
Ship where they already work
Streaming, voice, their language, their tools. If it isn't used, it isn't done.
Leave it understandable
Docs, tests and a handover the team can run without me.

Everything else

From 2022 onwards: computer vision, speech, retrieval and agents. Search for a tool or pick an area.

Search projects

26 projects

Status	Project	What it does	Stack	Link
Prototype	AI Solution Planner	Six agents turn a goal and its constraints into a structured solution blueprint	Phidata, OpenAI, Streamlit, Flask
Prototype	AI Teaching Agent Team	Agents that generate learning content, roadmaps and curated resources for a topic	Phidata, GPT-4o, Streamlit
Prototype	Blog Post Generator	Planner, writer and editor agents produce a post from a topic	CrewAI, GPT-4o, Streamlit, Flask
Prototype	OmniChat	One chat interface across model providers, with voice input	Anthropic, OpenAI, Streamlit
Experiment	MCQ Generator	Generates multiple-choice questions	Python, Jupyter	code ↗
Experiment	Sentiment classification	LSTM trained on 35M Amazon reviews, served in a web app	TensorFlow, Keras, Django	code ↗
Prototype	Smart Document Chat	Ask questions across a set of uploaded documents	LlamaIndex, Flask
Prototype	Smart Database Chat	Chat with a PostgreSQL database in plain language	LangChain, PostgreSQL, Flask
Prototype	Conversational SQL Generator	Turns a conversation into SQL through a small web API	LangChain, Gemini, OpenAI, Flask
Prototype	Realtime speech transcription	Streams microphone audio over WebSockets and returns live text	Sarvam STT, WebSockets, Python
Prototype	Speech Synthesis Tool	Text-to-speech through a web interface and API	Streamlit, Flask
Prototype	Video Transcription	Converts video and audio files to text	Transformers, Gradio, Flask
Prototype	Multilingual video dubbing	Transcribe, translate, re-voice and lip-sync a video	Whisper, XTTS, Wav2Lip	code ↗
Prototype	Story2Video	Story to scenes to images to narration	GPT-4, Stable Diffusion, MeloTTS
Prototype	YouTube Summarizer	Multilingual highlights from video transcripts	LangChain, Gemini, Groq
Prototype	Subtitle Translate	Transcribes and subtitles uploaded video in many languages	Speech recognition, MT	code ↗
Experiment	Speech emotion recognition	Seven emotions from pitch, volume and energy	CNN, audio features	code ↗
Internship	Docvision	Invoice field extraction at Chaitanya AI, mAP50 0.764	YOLOv8, OCR, Flask
Prototype	Online exam proctoring	Computer vision that flags cheating during online exams	Computer vision, Python
Prototype	Number plate detection	Detects plates in video and reads them with OCR	YOLO, OpenCV, EasyOCR
Experiment	Real-time object detection	80 classes from webcam or video	YOLOv8, OpenCV
Experiment	Face recognition demo	Recognises known faces in images and video	face_recognition, Flask
Experiment	BotaniScan	Detects 24 Indian medicinal plants	YOLOv8
Experiment	Retinal OCT classification	Eye-disease classification with transfer learning	ResNet50, Flask	code ↗
Prototype	Text-to-Interior	Generates interior designs, then detects the furniture in them	Stable Diffusion, YOLO, Flask
Prototype	Creative AI Image Maker	Image generation through a web interface and API	OpenAI, Streamlit, Flask

Now

BuildingUmmul Kitab and Haya, and my own agent systems on the Claude API.
ReadingHistory, philosophy, economics — notes will be published here.

If you're deploying AI to real customers, I'd like to hear about it.

mozakariyamansuri@gmail.com

I take AI out of the demo and into the hands of the people who use it.

Message Notification Router

An AI assistant for financial advisors

Asking company databases questions in plain language

Ummul Kitab

Haya

15Hisaab

Trivium

How I work with clients

Find the real problem

Prototype on their data

Ship where they already work

Leave it understandable

Now

If you're deploying AI to real customers, I'd like to hear about it.