Principal DeveloperNov 2024 — Present
Readable English
A reading tutor that listens — and keeps listening when the Wi-Fi drops.
- platforms, one on-device speech engine
- 3
- speech recognition, no server round-trip
- Offline
Jinesh Soni
Principal Developer at AppExert
Software that works after the demo.
Plenty of AI looks great on stage. I take it the last mile — onto phones, into classrooms with patchy Wi-Fi, into apps people use every day. 12 years of shipping mobile products taught me what production really means; now I architect the AI that runs inside them.
- AI architecture
- On-device models · pipelines
- AI in production
- Offline speech recognition
- Mobile foundation
- 12 yrs · 600K+ downloads
- platforms running on-device speech AI
- 3
- downloads across Bijak & Just Fresh, which I led
- 600K+
- years shipping production software
- 12+
- engineers led at Bijak
- 8
AI work
AI I’ve shipped — not just prototyped.
The hard part of AI isn’t the model. It’s everything around it: making it fast on a cheap phone, working offline, and turning a model’s guesses into something a person can trust.
In production · Readable English
On-device speech AI
sherpa-onnx with a memory-optimised local model, running fully offline on iOS, Android and the web — fast enough for word-by-word feedback.
Read the case studyIn production · Readable English
Turning model output into decisions
A text-matching algorithm that aligns noisy speech recognition with the expected passage to work out exactly which words a learner has read.
Read the case studyIn production · Readable English
Generative content pipeline
An AI pipeline that generates passages and practice material levelled to each learner’s Lexile and reading grade — the content a family of learning games runs on.
Read the case studyIn production · Bijak
ML for operations
TensorFlow models and Gen AI tooling that took repetitive manual work out of an agri-trade company’s internal processes.
Read the case study
Exploring now
Small, private models inside real products — agents and retrieval that keep working on the device.
AI agents · RAG · Local LLMs (Ollama) · Embeddings · Fine-tuning · LangChain
Selected work
Products where AI meets the real world.
Each case study covers the tension at the heart of the project, the decisions I made, and what they cost — from on-device speech AI to Flutter apps with 600K+ downloads.
Engineering LeadJul 2020 — Oct 2024
Bijak
Putting a trust-based, cash-heavy trade into an app — in eleven languages.
- downloads, incl. Just Fresh
- 600K+
- languages
- 11
Engineering LeadJul 2022 — Oct 2024
Just Fresh
A family of apps on one codebase — 80% of it shared — for produce that changes price every morning.
- downloads (within Bijak’s 600K+)
- 50K+
- code shared across the apps
- 80%
+ 20 more case studies & projects — CoinTopper, Wandergier, Roche Kibo, Studentdesk, open-source tools and more.
Browse the archiveHow I work
A few things I’ve learned the hard way.
01
The model is the easy part.
A speech model only became a reading tutor once we wrote the logic that decides which words were really read. The value of AI lives in the system around the model — data, latency, fallbacks and the judgement layer.
02
Use AI where it removes work.
The best AI I’ve shipped is invisible: a model that saves an ops team hours, or speech recognition that just works offline. Demos are easy; dependable is the job.
03
Build with AI, not just for it.
Claude Code, Cursor and GitHub Copilot are part of how I work every day — scaffolding, refactors, tests and reviews move at a pace that wasn’t possible a couple of years ago. The judgement stays human: I own the architecture, review every change, and let the tools do the typing.
04
Offline is a feature, not an edge case.
Traders in mandis and kids in classrooms don’t have perfect networks. I design for the bad connection first — local data, on-device models, graceful sync — and the good connection takes care of itself.
05
Architecture is for the team that maintains it.
I pick patterns a new hire can follow on day three. A boring, consistent structure across apps beats a clever one that only its author understands.
06
Measure before you argue.
Analytics, remote config and crash reports settle debates faster than meetings. Ship behind a flag, watch the numbers, then decide.
07
Make the path to production boring.
Automated builds, tests and store releases mean shipping is a non-event — so the team ships small and often instead of big and scary.
Experience
From Android apps to AI systems.
Nov 2024 — Present
Remote
Principal Developer, AppExert
Remote engineering teams for product companies.
Leading architecture and delivery of Readable English — a cross-platform Flutter reading app with on-device speech recognition and AI-generated learning content.
Show highlightsHide highlights
- Architected and led the Flutter app across iOS, Android and Web.
- Shipped fully on-device speech recognition with sherpa-onnx on all three platforms, using a memory-optimised local model for fast STT.
- Developed a text-matching algorithm that accurately detects which words of a passage have been spoken, powering real-time reading feedback.
- Built multiple games on this engine for decoding, vocabulary building and fluency (Falling Words, Vocab Vault, Space Race).
- Set up an AI pipeline that generates app content matched to each learner’s Lexile and reading grade level.
- Wired up Firebase Remote Config, Analytics and Crashlytics, and shipped in-app subscriptions.
Jul 2020 — Oct 2024
Gurugram, India · Remote
Engineering Lead, Bijak (Krishiacharya Technologies)
India’s B2B agri-trade marketplace for mandi traders.
Involved from day one — led Bijak’s move from native to Flutter and a team of 8 building the Bijak app and the Just Fresh family of apps.
Show highlightsHide highlights
- Led the migration from native Android to Flutter while continuing to ship features.
- Built B2B (Bijak) and D2C (Just Fresh) products end-to-end — 600K+ downloads combined.
- Chose Flutter 3 with GetX and MVVM to build fast with a consistent architecture.
- Led a team of 8 engineers, coordinating with design, product and marketing.
- Built the Just Fresh apps — customer, delivery & logistics and warehouse admin — with 80% shared code.
- Automated build, test and release with CI/CD (Codemagic).
- Integrated analytics and engagement tooling (Segment, MoEngage, Branch) to measure and optimise funnels.
- Trained TensorFlow models and introduced Gen AI tooling to cut manual work in internal processes.
Dec 2018 — Jul 2020
Ahmedabad, India
Team Lead — Android & Flutter, Thinkwik
Product studio building apps for clients worldwide.
Led the mobile team delivering client apps across healthcare, travel, sports and agriculture.
Show highlightsHide highlights
- Led a mobile team through the full SDLC to deliver production-grade apps.
- Gathered user feedback and app-acceptance research to make new and existing apps easier to use.
- Championed a Kotlin-only codebase with MVVM architecture.
- Shipped apps including Roche Kibo, Wandergier, Footy Addicts and Agrodei.
Apr 2016 — Dec 2018
Ahmedabad, India
Software Engineer, CodeCrunch TechLabs
Ahmedabad product lab behind Studentdesk and CoinTopper.
Built the company’s flagship products Studentdesk.in and CoinTopper, plus a string of client apps.
Show highlightsHide highlights
- Developed Studentdesk.in and the CoinTopper real-time crypto tracker.
- Used test-driven development to refine products.
- Scaled apps to handle 10,000+ daily active users.
- Applied Material Design and a hybrid Kotlin/Java MVP architecture.
Jul 2014 — Mar 2016
Ahmedabad, India
Android Developer, eSwasthya Mediserve
Online pharmacy and wellness delivery.
First engineering role — designed and built the pharmacy-delivery app and its order-processing backend.
Show highlightsHide highlights
- Drafted, designed and developed the Android app.
- Designed the backend system for order processing.
About
I started by breaking my own phone. Then I learned to build.
Around 2012, in university, I was modding Android ROMs to squeeze more performance out of my phone — and writing about it on my tech blog, TechCure. It grew to 400K+ visits and taught me Android and Java inside out. In 2014 I started building apps professionally and never looked back.
Since then I’ve worked across healthcare, crypto, travel, agri-tech and ed-tech — first as an Android engineer, then leading mobile teams, and today as Principal Developer at AppExert, architecting an AI-powered reading app with on-device speech recognition.
The thread through all of it: apps for people who don’t have perfect conditions. Traders on budget phones in busy mandis. Kids reading aloud in classrooms with flaky Wi-Fi. I like the constraints — they make the engineering honest.
And increasingly, the work is AI. Not demos — AI that has to hold up in the same messy conditions. At Readable English that meant running speech recognition entirely on-device with sherpa-onnx across iOS, Android and the web, writing the matching algorithm that turns a recogniser’s guesses into word-accurate reading feedback, and building an AI pipeline that generates learning content levelled to each reader’s Lexile and grade. At Bijak it was TensorFlow models and Gen AI tooling that quietly took manual work off the operations team.
Next, I’m exploring where agents, RAG and local LLMs fit into mobile products — small models that run on the device, respect privacy, and keep working when the network doesn’t.
- Education
- B.E. Computer Science, Gujarat Technological University
- Languages
- English, Hindi
Toolbox
The stack I reach for.
AI & ML
Generative AI · LLMs · RAG (Retrieval-Augmented Generation) · AI agents · LangChain · Local LLMs (Ollama) · Embeddings · Fine-tuning · On-device speech recognition (sherpa-onnx) · AI content pipelines · Machine Learning · TensorFlow
Mobile
Flutter · Dart · Android · Kotlin · Java · iOS · Flutter Web · Jetpack
Architecture & state
MVVM · MVP · BLoC · GetX · Riverpod · Provider · SOLID · Offline-first · Dagger2 · RxKotlin
Data & backend
Firebase · REST · GraphQL · WebSockets · MQTT · SQLite · Hive · Realm · MySQL · Node.js · Flask · Python
Ship & measure
AI-assisted development (Claude Code, Cursor, Copilot) · CI/CD (Codemagic) · A/B testing · Remote Config · Crashlytics · Segment · MoEngage · Branch.io · Docker · AWS · GCP
Leadership
Team & people management · Mentoring · Code review · Agile / Scrum · Cross-functional delivery
Contact
Building something with AI?
Taking AI from prototype to production, putting models on the device, architecture reviews or leadership roles — I’d love to hear what you’re building.