Mohammad Zaki Jariwala
AI Product Manager · Technical Lead Infrastructure + AI: shipping products in the gaps Engineer who builds products and wants to do it full-time
Three years keeping critical infrastructure running. Now building AI products full-time. whatever I'm building this month I find it hard to stop building things
Building a content automation pipeline that turns YouTube transcripts into publication-ready articles across 12 niches. Studying for GCP Professional Cloud Architect. Reading about product discovery frameworks and what makes an AI product actually useful once the novelty wears off.
About
Systems Engineer at TCS, deployed at State Bank of India, keeping mission-critical banking infrastructure running at 99.999% uptime for 30,000+ users. The day job built the technical credibility: three years leading a team of five, managing 150+ servers, cutting RTO by 50%. The nights built the product experience: an offline supplication app with 22,000 records and sub-10ms search, a running app with 50+ alpha users, a marine equipment digital ecosystem with AI-powered B2B outreach, and a content automation pipeline. I know what engineers are actually talking about. I know what it takes to ship something real.
I run infrastructure for a Tier-1 Indian bank by day and build AI automation pipelines, PWAs, and embedded devices by night. My stack spans Linux/RHEL server fleets, Python ETL and LLM orchestration, Flutter mobile, Astro JAMstack, and ESP32 firmware. Usually on free-tier infrastructure and always with a bias toward shipping over theorising. Currently deep into GenAI workflow design and thinking hard about what product management looks like when the core technology is still moving.
I'm an engineer based in Navi Mumbai who finds it hard to stop building things. My day job is keeping banking systems alive for tens of thousands of people. That part I'm good at. But it's not the part that keeps me up.
Outside that, I've built an offline Islamic supplication app used by a small community of Muslims worldwide, a running app with ghost routes and crowd radar, a pocket e-ink reader from scratch, and a marketing automation system for my family's marine parts business. None of these were assigned to me. I built them because the alternatives were bad and I thought I could do better.
I also spent a few months coaching engineers on how to talk about their work, which turned out to be more technical than it sounds. Understanding what makes a story land is not that different from understanding what makes a product land.
Right now I'm building a content automation pipeline and thinking seriously about moving into product management. Not because it sounds good on a resume, but because the part of every project I've enjoyed most is always the same: deciding what to build and why.
Experience
Tata Consultancy Services
Deployed at State Bank of India (SBI GITC)
Systems Engineer, IDM & Infrastructure
For my day job I keep a major bank's systems running for tens of thousands of people. The kind of work where success means nobody ever notices it.
- Led an operations team delivering 99.999% availability for Tier-1 banking applications across 30,000+ branch users: setting team priorities, mentoring 5 engineers, and managing escalations end-to-end.
- Drove infrastructure modernisation, evaluated build-vs-buy tradeoffs, secured stakeholder alignment, and led adoption of on-premise container-based workflows within compliance boundaries. Translated engineering constraints into a prioritised delivery roadmap for leadership.
- Architected and executed Disaster Recovery runbooks in cross-functional coordination with application and database teams, reducing Recovery Time Objective (RTO) by 50%.
- Operated IBM Security Identity Manager (ISIM) on RHEL 8, managing user provisioning, access governance, and policy enforcement across 150+ application servers in an on-premise Tier-1 banking datacenter maintaining 99.999% platform SLA via active-passive HA clustering and automated health monitoring.
- Led on-premise containerisation of legacy IDM workflows: evaluated Podman vs Docker within RBI/SEBI compliance boundaries, defined container topology and orchestration approach, secured sign-off from security and compliance stakeholders.
- Authored cross-team DR runbooks covering IDM primary/replica switchover, application server restart sequences, and database consistency validation, reducing RTO from ~4 hours to under 2 hours (50% reduction) through structured tabletop drills and live execution.
- Automated recurring ops tasks (patch compliance checks, certificate rotation tracking, capacity utilisation reports) in Python and Bash, eliminating ~8 hours/week of manual overhead across the team.
- The day job means keeping banking infrastructure alive for tens of thousands of people. Success is invisible: if nobody notices, everything went right.
Independent Coaching Practice
Technical interview readiness for engineering and management candidates
Founder & Technical Communication Coach
For a while I ran a small coaching practice helping engineers talk about their work in a way that actually lands with people who aren't engineers.
- Founded and scaled a coaching practice from zero, handling client acquisition, curriculum design, and end-to-end delivery for 50+ engineering and management professionals.
- Designed a structured curriculum covering STAR-method framing, stakeholder communication, and senior-level interview strategy, iterating based on client outcomes and session feedback.
- Developed repeatable frameworks for technical storytelling, enabling clients to translate complex engineering work into language accessible to non-technical decision-makers.
- Turned out that teaching engineers to explain themselves clearly is a harder problem than it looks, and more satisfying to solve than I expected.
Projects
Zamaan Marine Digital Ecosystem
Full digital ecosystem for a family marine parts business: storefront, cold email automation, and analytics at zero hosting cost.
- Identified the digital gap through direct conversations with 200+ B2B prospects, validating channels, value proposition, and pricing signals before writing a line of code.
- Defined and prioritised a zero-budget digital stack: JAMstack storefront for discoverability, eBay API automation for listings (~60% faster time-to-market), and an AI-powered cold-email funnel for B2B acquisition.
- Instrumented product analytics via PostHog, tracking the full funnel from storefront visit to eBay conversion. Used data to reprioritise listing categories and outreach targeting.
- Shipped an AI cold-email agent achieving 90%+ inbox delivery across campaign cycles.
- Cold-email agent: Claude API prompt templates per industry vertical → Python orchestrator → Brevo SMTP with sender rotation and bounce handling → delivery rate tracking. 90%+ inbox rate across campaign cycles.
- eBay Browse API integration: Python ingestion pipeline parses product data, maps to eBay taxonomy, and pushes listings in batch. ~60% reduction in per-SKU listing time vs. manual entry.
- PostHog JS SDK on a fully static Astro/Cloudflare Pages frontend: custom event schema tracking browse → click → eBay redirect funnel, session replay configured to exclude contact fields.
- Zero-persistent-backend architecture: Astro SSG → Cloudflare Pages CDN, GitHub Actions CI, all serverless. Total recurring cost: $0.
- My family's business had no real digital presence. I built the whole thing — website, automated email campaigns, product listings — in spare time, at zero cost.
My family started a business selling refurbished marine and industrial parts, mostly through eBay. The digital side was an afterthought. I stepped in to build it properly, which meant treating a small family operation with the same engineering rigour I'd apply at work. The constraint of zero budget made it more interesting, not less.
Content Automation Pipeline
YouTube Short to 6-platform publication-ready content pack in under 3 minutes: Medium article (2000+ words) plus Twitter/X, LinkedIn, Bluesky, Reddit, and Threads variants.
- Identified that content-to-publication time was dominated by mechanical, repeatable steps. Designed and shipped a pipeline automating transcript ingestion, LLM transformation, and multi-platform output across 6 platforms.
- Reduced time-to-draft by 85%; pipeline runs on a GitHub Actions workflow_dispatch trigger. Open GitHub app, enter URL and niche, green checkmark in under 3 minutes, output committed to repo.
- Built a quality gate layer ('Failure First') that detects safe/generic LLM output and forces re-generation until prose passes specificity checks, addressing the core failure mode of LLM-generated content.
- 12 niche-specific prompt frameworks adapt tone, structure, and framing per domain without modifying the core pipeline.
- Two-call Gemini 1.5 Flash architecture: Call 1 generates the full Medium article (2000+ word hard floor enforced in prompts/system.md); Call 2 derives all 5 social platform variants from the finished article body, not the raw transcript, ensuring coherence across outputs.
- Failure First framework: heuristic validators scan for hedging language, filler phrases, and passive-voice saturation; triggered gates resubmit with escalating specificity constraints until output clears all checks.
- CI mode: _IS_CI = bool(os.getenv('CI')) — GitHub Actions sets this automatically, eliminating all input() calls that would hang the workflow runner.
- Bluesky auto-posting via atproto library. Thread constructed from generated posts, published live if credentials are set. Medium pushes as draft only via markdown2medium. yt-dlp subtitle extraction with manual transcript fallback for Shorts lacking caption tracks.
- I was writing articles and the process was the same mechanical steps every time. I automated those steps. The only interesting problem was making the output not sound like a robot, which became a framework for detecting when AI writing is being safe and forcing it to be specific instead.
I was writing AWS certification articles for Medium and realised the research-to-draft process was almost entirely mechanical. I built a pipeline to handle the mechanical part so the human part — judgment, voice, editing — could actually matter. The Failure First framework came from noticing that AI-generated content fails the same way every time: it starts safe and stays safe. I built rules specifically to break that pattern.
Zamaan Automation
Data-driven SEO layer for a second-hand industrial automation parts eBay store. Formal experiment with null hypothesis, UTM attribution, and a 90-day decision gate built into the architecture.
- Framed a formal experiment before writing a line of code: defined null and alternative hypotheses, a quantitative 90-day success threshold (50+ UTM-attributed eBay clicks/month), and explicit shutdown criteria if the threshold isn't met.
- Built the complete SEO discovery layer: Astro 4 static site, automated daily eBay listing sync via Browse API, UTM-attributed CTAs across 6 product categories, Pagefind full-text search, and PostHog analytics. Total monthly cost: £0.
- Shipped 5 SEO blog posts at launch (PLC buying guide, OEM vs refurbished cost analysis, VFD guide, HMI panel guide, Siemens S7-1200 guide) to seed organic traffic before the 90-day measurement window opens.
- TypeScript eBay Browse API integration (src/lib/ebay.ts): OAuth client_credentials token flow → paginated item search (200/page) with filter=sellers:{id} → category and condition derived programmatically from eBay taxonomy strings → results cached per build.
- GitHub Actions daily sync (02:00 UTC): npm ci → astro build && npx pagefind --source dist with eBay env vars injected → Cloudflare Pages deploy via cloudflare/pages-action.
- UTM attribution fully typed in src/lib/utm.ts. buildEbayUrl() and buildStoreUrl() enforce utm_source=zamaanautomation, utm_medium=organic, utm_campaign={pageType}, utm_content={category} on every outbound link.
- Pagefind: static full-text search index built at compile time, runs in the browser, zero server, zero API cost. PostHog event schema tracks BUY_ON_EBAY_CLICKED and CONTACT_FORM_SUBMITTED for the full funnel from discovery to eBay redirect.
- The experiment question is: does having a website actually help sell more on eBay? I built the infrastructure to test it properly, with a real null hypothesis and a real deadline for the decision. It either reaches 50 clicks a month by Month 3 or the content investment stops.
The family marine business already had a digital presence and email automation. The open question was whether a dedicated SEO site would actually drive eBay sales. I ran it as a formal experiment: null hypothesis, 90-day measurement window, explicit shutdown criteria. The decision to stop would be as evidence-driven as the decision to build. I wrote the shutdown criteria before I wrote the first blog post.
run.to
Zero-build PWA that generates closed-loop running routes to a precise target distance or time. Hexagonal geometric projection snapped to real pedestrian paths via OSRM, with per-km splits and GPX export.
- Identified a real gap: existing route planners require manual path drawing. Built a generator that automatically produces closed-loop routes converging on a target distance or time.
- Shipped as an installable PWA. Works offline after first load, opens from home screen with no app store required.
- Two input modes (distance and time), per-km splits, GPX export for Garmin/Strava compatibility, and a Google Maps deep-link for walkers who want turn-by-turn navigation.
- Hexagonal geometric projection engine (routeEngine.js): calculates a 6-sided polygon with a random rotational offset around the start point, then snaps all 6 waypoints to pedestrian routes via OSRM /route/v1/foot/. Random offset generates a different route orientation on every run.
- Fuzzy convergence: radiusFactor is adjusted each iteration based on the ratio of OSRM-returned distance to target distance. Loop continues until within tolerance or max iterations reached.
- Haversine fallback: when OSRM is unreachable, routeEngine.js computes geometric distance using the Haversine formula and renders the raw polygon as 'Simulation Mode'. App stays usable without the API.
- Zero-build stack: Leaflet 1.9.4 + Tailwind CSS + Lucide icons all loaded from CDN. No npm, no bundler. Service worker caches app shell for offline use. Time mode: user input in minutes → converted at 5.5 min/km → same route engine runs.
- I wanted a route that came back to where I started and hit a target distance. Every app I tried needed me to draw it myself. I built one that figures out the geometry: picks a hexagon, measures how far off it is, adjusts until it's close.
I wanted a route that came back to where I started and hit a target distance. Every app I tried needed me to draw the route myself. I built one that figures out the geometry: it picks a hexagon shape, tries it, measures how far off it is, and adjusts until it's close. That feedback loop is exactly how I run.
Path of Supplication
Offline-first PWA delivering 22,000+ lines of multilingual Islamic liturgical content across 5 categories: sub-10ms search, 80MB offline payload, zero hosting cost.
- Identified a gap in available supplication apps. Existing options were ad-heavy, required constant connectivity, or lacked multilingual support. Validated the problem before building.
- Structured 22,000+ liturgical records across 5 content categories (Qur'an, Namaz, Dua, Ziyarat, A'maal) into an indexed export pipeline achieving sub-10ms client-side search.
- Shipped as an offline-first PWA with service worker caching, delivering 80MB of multilingual content (Arabic, Urdu, English) at zero hosting cost.
- Ran a structured beta with 40+ testers across iOS Safari, Android Chrome, and desktop, triaging feedback and shipping iterative improvements across service worker versions up to v9.
- Zero-build architecture: pure HTML/CSS/JS, no framework, no npm. Service worker (sw.js, v9) pre-caches the app shell; 80MB content payload is lazy-loaded and persisted in Cache API indefinitely.
- Content pipeline: ron.db (87MB SQLite, authoring tool only, never committed) → scripts/export.py → 2,704 JSON files in docs/data/lines/ + 4 index files (categories.json, nav.json, search.json, audio.json). Browser fetches pre-exported JSON, no SQLite runtime in browser.
- Deep Search: search.json is a pre-built multilingual full-text index (Arabic/Urdu/English/Roman Urdu) scanned in a single vanilla JS pass, no query engine, no dependency, sub-10ms on mobile.
- Cache invalidation: bump VERSION constant in sw.js to force all clients to flush stale caches. Active design work: Mockups/ directory contains 10+ screen specs used for iterative UI refinement.
- The apps that existed for Islamic duas and ziyarat were slow, plastered with ads, or useless offline. I built one that works on any device, in your language, with no internet after the first load.
The existing apps for Islamic supplications were clunky, ad-heavy, or required a constant internet connection. For something as personal as duas and ziyarat, that friction felt wrong. I wanted something that worked offline, in your language, on any device, with no ads and no backend. A big part of the work was understanding the data well enough to build proper search across five content types in three scripts.
Ian Xiaohei Illustrations
Open-source Claude Code skill for generating 16:9 hand-drawn article illustrations: 9 culturally-grounded character variants, 4 image generation providers with auto-detection, ported and expanded from a Chinese Codex skill.
- Identified an open-source gap: the original Chinese-language Codex skill produced no actual images and had a single character. Ported to English, added real multi-provider image generation, and expanded to a 9-character cultural variant system.
- Designed a character system grounded in distinct art traditions (ukiyo-e, Madhubani, WPA poster, Arabic calligraphy, Adinkra/kente) to avoid cultural caricature. Philosophy documented and published alongside the code.
- Ran two rounds of structured multi-agent review (4 parallel agents per round: Newcomer, Engineer, Growth Strategist, Docs Editor) before publishing. All findings resolved.
- Published 14 example illustrations in the repo demonstrating the style across varied cognitive structures.
- SKILL.md entrypoint for Claude Code: reads settings.json for default character, scans project context if no article is provided, identifies cognitive structure, selects character, calls scripts/generate_image.py.
- Provider auto-detection in generate_image.py: checks env vars in order (GEMINI_API_KEY → OPENAI_API_KEY → STABILITY_API_KEY), selects first available provider, generates a 16:9 PNG. Provider aliases handled (gemini → nanobanana, openai → dalle, sd → stability).
- DALL-E routing: uses gpt-image-1 (1536x1024) as primary; falls back to dall-e-3 (1792x1024) automatically. stdlib Python for Gemini, DALL-E, and Imagen paths, only Stability AI requires pip install requests.
- install.sh is idempotent: copies skill files to ~/.claude/skills/, reads existing settings.json and merges rather than overwriting. User's default character survives upgrades. Multi-platform: Claude Code (SKILL.md), Codex, Gemini CLI, Hermes, Antigravity agent configs all included.
- I found a Chinese skill for drawing article illustrations, ported it to English, and kept going, adding real image generation and expanding one character into nine. The interesting constraint was figuring out how to make each character feel like it belongs to its cultural heritage rather than just borrowing the aesthetic.
I found a Chinese-language Codex skill for illustrating articles and wanted to use it, but it didn't generate actual images and only had one character. I ported it to English, added real multi-provider image generation, and expanded one character into nine. The interesting constraint was making each character feel genuinely grounded in its cultural heritage — ukiyo-e, Madhubani, WPA posters, Arabic calligraphy, Adinkra — rather than just borrowing the aesthetic.
GidsTek Website
Full Astro 6 replacement for an IT hardware reseller's WordPress site: 14.54s cold load to sub-1s, CLS 1.0 to 0, all 5 missing security headers added. Delivered as a barter against a hardware purchase.
- Ran a full technical audit before quoting: 14.54s cold load, CLS of 1.0 (maximum possible), B-grade SSL, 5 missing security headers, and indexed WordPress demo pages still in Google's index.
- Scoped and delivered a complete replacement site in Astro 6: hardware catalog across 6 categories, WhatsApp inquiry CTAs, all performance and security issues resolved, as a barter against a product purchase.
- Agency equivalent in Mumbai: Rs 25,000–60,000. Delivered as leverage in a price negotiation, not as a paid engagement.
- Astro 6.4.4 static site with TypeScript throughout: src/data/hardware.ts and src/data/services.ts define typed catalog data across 6 hardware categories (laptops, servers, networking, firewall, UPS, CCTV).
- public/_headers: adds all 5 security headers missing from the original site (X-Frame-Options, X-Content-Type-Options, Referrer-Policy, Permissions-Policy, Content-Security-Policy), fixing the Snyk 'Grade capped at A' finding.
- public/_redirects: maps all legacy WordPress URLs to current equivalents, preserving any inbound links after migration. CSS-only hamburger nav, no JavaScript toggle, no event listeners.
- I was about to buy a laptop and noticed the shop's website was unusably slow. I ran an audit, built a replacement, and offered to trade it for a discount. The whole negotiation ran on technical evidence: a Lighthouse report, a Snyk scan, and a cold-load comparison.
I was buying a laptop from an IT hardware shop and noticed their website was embarrassingly slow. I ran a full technical audit and built a complete replacement in Astro 6. Then used the audit as a negotiating tool: trade the site rebuild for a discount on the hardware. The whole negotiation ran on technical evidence.
ESP-Pocket Reader
A pocket e-ink reading device built on ESP32-S3: custom firmware, SQLite content storage, and a Python EPUB ingestion pipeline.
- Designed and built a working hardware product end-to-end: from EPUB ingestion pipeline to custom embedded firmware with stable e-ink display rendering.
- Architected a custom pocket-sized e-ink reading device using an ESP32-S3 microcontroller and a 3.7-inch E-Paper display, achieving a functional firmware prototype with stable display rendering.
- Developed embedded C++ firmware (ESP-IDF/PlatformIO) integrating SD card storage and SQLite for robust content management, storing and paginating 50+ EPUB-converted documents in testing.
- Engineered a Python EPUB-to-SQLite conversion pipeline, reducing per-book ingestion time to under 30 seconds for titles up to 500 pages.
- Reading on a phone is miserable. I built my own pocket e-ink reader from scratch. Designing every layer yourself, from the EPUB parser to the display driver, turns out to be more satisfying than buying one.
Reading on a phone destroys your eyes and your battery life. E-readers exist but they're expensive, locked down, and designed around ecosystems I don't want to be in. I wanted a device that did exactly one thing: display text clearly, at almost nothing to run. Building the firmware myself meant understanding every layer of the stack, from EPUB parsing down to the display driver.
cos/os
A markdown-based project operating system: per-project status files, Telegram idle alerts, Notion sync via GitHub Actions, and a public /now page.
- Built a lightweight project OS to maintain visibility across 8 active builds: each project has a markdown status file tracking current phase, focus, and next action.
- Telegram bot (Cloudflare Worker) sends stale-project alerts when a project file hasn't been updated in over 7 days. No context loss across parallel builds.
- Markdown-first architecture: per-project .md files in cos-os/ committed alongside the portfolio repo. GitHub Actions sync to Notion on push. Zero external dependencies for the core tracking system.
- Cloudflare Worker handles Telegram webhook: parses incoming messages, reads project file state, sends structured status summaries. Stale-check cron triggers daily at 08:00 UTC.
- Public /now page parses cos-os/projects/portfolio.md at Astro build time via node:fs — no intermediate JSON, no sync script, no extra workflow. Page updates on every push to main.
- I had too many projects going at once and kept forgetting what was stuck. I built a simple system where each project has a file that says what's happening now and what's next. A bot tells me when I've been ignoring something. The lab page is just the public version of that.
I had too many things in motion at once and kept losing track of what was blocked, what was moving, and what hadn't been touched in weeks. I built a simple system: each project gets a markdown file with its current status, what I'm doing now, and what's next. A bot pings me when something's gone stale. The public lab page is the same data, sanitised. It felt more honest than pretending all projects are always finished.
Personal Remote Dev Server
A persistent cloud Linux dev environment on Oracle Free Tier: full-stack workflows from any device, zero cost, 99%+ uptime over 18 months.
- Set up and maintained a persistent cloud development environment with 99%+ uptime over 18 months, enabling full-stack development workflows from any device at zero recurring cost.
- Designed and configured a fully persistent remote Linux development environment on Oracle Cloud Free Tier: production-equivalent compute at zero recurring cost.
- Optimised for smartphone and laptop access via Zellij terminal multiplexing and Termius SSH, enabling full-stack development workflows from any device.
- Maintained 99%+ uptime over 18+ months of continuous use, with automated session persistence eliminating context loss between work sessions.
- I got tired of long-running builds dying when my laptop went to sleep. I set up a server in the cloud that's always on, which means I can write code from my phone when I'm away from a desk.
My personal projects involve a lot of long-running processes — builds, pipelines, SSH tunnels — that don't play well with a laptop that goes to sleep. I also wanted to be able to develop from my phone when away from a desk. Oracle Cloud's free tier is genuinely powerful if you configure it right, and Zellij with persistent sessions solves the context-loss problem that makes remote development frustrating.
Skills
Certifications
AWS Certified Solutions Architect – Associate (SAA-C03)
Amazon Web Services
Authored a 3-part Medium article series on SAA-C03 aimed at non-technical candidates.
Google GenAI Leader
Google Cloud
Validates applied GenAI product thinking and responsible AI leadership — covering model evaluation, prompt design, and AI product strategy across the Google ecosystem.
Professional Cloud Architect
Google Cloud
Actively studying — expected completion mid-2026.
Contact
Let's talk product Let's build something Say hello
Open to AI PM, product, and innovation roles. Technical background with real 0-to-1 product experience. Based in Mumbai, open to relocation. Always interested in interesting problems. Best reached over email. Say hello. I reply.
Jariwalazaki@gmail.com