ShivankMathur
Building high-performance products at the intersection of great engineering and thoughtful design. 7+ years across the full stack — from React micro-optimizations to backend architecture.
Where I've Built
7+ years shipping across startups, agencies, and my own products.
New Relic
Senior Software Engineer
Working with the UX platform team to develop micro front-end applications at scale. Leading a pod of 6 engineers on the Maps team — building graph visualizations that surface relationships and critical impact areas across 600+ monitored application entities.
- Led a 6-engineer pod delivering New Relic Maps — a graph-based visualization of 600+ monitored application entities and their service dependencies
- Delivered 8 micro front-end applications in collaboration with the UX platform team, improving platform responsiveness across the product suite
- Improved identification of critical impact areas via advanced graph-based data representations for reliability engineering teams
Zeda.io
Senior Product Engineer
4 years building the core product at Zeda.io — a product intelligence platform. Owned the frontend architecture, integrated AI-driven features including feedback analysis with embeddings and clustering, and mentored junior engineers across the team.
- Architected AI-powered feedback intelligence — embedding-based clustering of 50K+ feedback items with ~87% accuracy, reducing manual triage time by 70%
- Integrated LLM-driven features using LangChain to create predictive, adaptive user experiences across the product
- Led design, development, and optimisation of responsive web interfaces — implemented best practices for performance, accessibility, and scalability
- Mentored junior developers; fostered a culture of innovation and continuous learning across the frontend team
Check Is It AI
Founder & Lead Engineer
Built and launched checkisitai.com — an AI media forensics tool that analyses images, video, and audio for signs of AI generation. Combines Claude's vision API with client-side Error Level Analysis, EXIF deep-dive, and a structured multi-signal verdict system.
- Designed a structured JSON schema for Claude responses covering AI probability, forensic signals, bounding box coordinates, and a full technical profile per upload
- Implemented client-side ELA using the Canvas API — zero server cost, runs in under 200ms, gives users a visual forensic layer before the model responds
- Atomic credit deduction via Postgres stored procedure — SELECT FOR UPDATE prevents race conditions on concurrent requests hitting the same user balance
- Live at checkisitai.com with Supabase Auth, per-user storage, and full credit transaction audit trail
LiquidHQ
Founder & Lead Engineer
Building a multi-service SaaS platform that lets Shopify merchants generate production-ready Liquid sections from natural language prompts. Microservices architecture across Vercel and Railway — embedded Shopify app, standalone AI pipeline, background job workers, and a shared Supabase database with row-level security.
- Architected a 7-repo microservices system: Shopify embedded app (Next.js), AI generation service (Fastify), background workers (BullMQ), database migrations (Supabase), marketing site, pixel SDK, and pixel relay worker
- Built an SSE streaming pipeline from Anthropic Claude through Fastify to the embedded app — merchants see Liquid code generate in real-time with syntax highlighting via CodeMirror
- Designed a versioned system prompt with few-shot examples that generates valid Shopify Liquid sections with schema, responsive CSS, and sensible defaults — validated server-side with liquidjs before returning
- Implemented per-shop rate limiting (Redis), tiered billing hooks (Shopify Billing API), and RLS-first database design where every table is scoped to shop_id
Caissa Labs
Founder & Lead Engineer
Built a chess analytics platform from 0 to production, solo. Real-time game analysis, pattern detection across hundreds of games, and personalised improvement insights powered by a batch API architecture designed for cost efficiency.
- Designed a batch API pipeline that processes games in bulk — 10× cheaper than per-request analysis calls
- Pattern detection engine with ~90% accuracy using positional hashing and clustering across opening, middlegame, and endgame phases
- Built the full stack independently: React board UI, Node.js API, PostgreSQL, deployed on Vercel
Publicis Sapient
Associate Technology
Contributed to financial sector projects at a global digital transformation agency. Built data visualisation tools, dashboards, and full-stack solutions for enterprise clients — collaborating with international teams across time zones.
- Developed full-stack solutions with React for interactive financial data interfaces serving enterprise clients
- Created data visualisation tools and dashboards presenting complex financial insights effectively to stakeholders
- Collaborated with international teams to deliver robust, client-focused solutions tailored to specific business needs
- Contributed to scalable architecture design for data-driven, user-friendly applications
Myaha Living
Full-Stack Developer
Designed and built a fully custom Shopify Plus store for a D2C home decor brand — custom Liquid theme, React-powered interactive sections, Checkout UI Extensions, and Shopify API integrations for inventory and fulfilment. Full ownership from brand strategy through deployment.
- Built a custom Shopify Plus theme from scratch — no marketplace template, full control over render path and Lighthouse score
- Sub-2s LCP on product pages — 35% better than D2C industry benchmark
- Integrated Shopify APIs, logistics webhooks, and Checkout Extensions in a streamlined purchase flow
- Launched in 6 weeks, 0 to first sale on day one
Case Studies
Selected projects with the full engineering story — problem, decisions, and outcomes.
Check Is It AI
FeaturedAI Media Forensics Detector
Problem
With AI-generated images, video, and audio becoming indistinguishable from real media, creators, journalists, and brands needed a tool that could forensically analyse uploaded files and return a credible, explainable verdict — not just a single confidence score with no reasoning behind it.
Approach
- Built a full-stack Next.js app with a secure server-side API route that sends media to Claude's vision API — keeping the API key off the client entirely
- Designed a structured JSON schema for Claude responses covering AI probability, verdict, 4–6 forensic signals, 3–5 analysis pointers, bounding box coordinates for suspicious regions, and a full technical profile
- Implemented client-side Error Level Analysis (ELA) using the Canvas API — re-compresses the image at a known quality and amplifies pixel-level differences to surface tampered or synthetically generated regions without any server round-trip
- Added EXIF metadata deep-dive using the exifr library, flagging missing standard fields, AI software signatures, and timestamp inconsistencies that suggest post-processing
- Built a credit system with atomic Postgres-level deduction via a stored procedure — preventing race conditions when concurrent requests hit the same user balance
- Integrated Supabase Auth for sign-in/sign-up, Supabase Storage for per-user media uploads, and a full audit trail of credit transactions
Engineering Notes
Prompt engineering to prevent false positives
Early versions flagged professional product photography as AI-generated because Claude over-weighted filename patterns (e.g. 'Gemini-14324.jpg' scored 85% suspicion from the name alone). Rewrote the system prompt to cap filename evidence at 5–10 points, explicitly call out that professional DOF and lighting are not AI signals, and instruct the model to treat human hands and environmental imperfections as strong authenticity indicators. False positive rate dropped significantly.
Atomic credit deduction
A naive read-then-update credit check has a race condition — two concurrent requests can both pass the balance check before either deducts. Solved with a Postgres stored function that does SELECT FOR UPDATE, checks balance, deducts, and inserts the transaction row in a single atomic operation. No application-level locking needed.
Client-side ELA with no server cost
Error Level Analysis is computationally cheap but was traditionally a server-side operation. Moved it entirely to the browser using the Canvas API — draw the original, re-compress to JPEG at a known quality, diff the pixel arrays, amplify by 20×. Runs in under 200ms on most images, adds zero API cost, and gives users a visual forensic layer before Claude even responds.
Unicode crash from btoa on emoji
The shareable URL feature encoded results as btoa(JSON.stringify(result)) — which throws on any character outside Latin1. Claude's responses include emoji in pointer icons. Fixed by replacing btoa/atob entirely with encodeURIComponent/decodeURIComponent, which handles the full Unicode range natively.
Tradeoffs Made
- Claude over GPT-4o: ~20–30% higher per-call cost but significantly more reliable structured JSON output — critical when the entire UI renders from parsed model responses with no fallback UI
- Metadata-only analysis for video and audio: Claude can't process binary video/audio, so forensics for those types is probabilistic based on filename and format patterns — the system explicitly warns users and caps confidence at Medium/Low for non-image files
- Credit system over subscription: lower friction for occasional users, and per-analysis pricing maps directly to the variable Claude API cost rather than absorbing it into a flat fee
Outcomes
Stack
How I Think & Build
The mental models and first principles that guide every technical decision I make.
Measure Before Optimising
Real user metrics first — Core Web Vitals, profiler traces, network waterfalls. Speculative performance work is a distraction. Every optimisation I've shipped has started with a specific number I wanted to move.
Batch Over Sequential
One batch call replacing 100 sequential requests isn't just faster — it changes what's architecturally possible. Caissa Labs went from 3s-per-game to 50-games-in-4s this way. I apply this pattern wherever I see sequential fan-out.
State Has Layers
URL state, server state, and UI state have different caching, invalidation, and ownership needs. Conflating them is where most frontend complexity bugs come from. I treat each layer as a separate concern with its own tool.
Product Tradeoffs Are Engineering Decisions
Choosing 87% clustering accuracy over 95% was an engineering call — I computed the marginal cost of the last 8 points and it didn't justify the compute. Good engineers surface these tradeoffs; great ones frame them in product terms.
Ship Small, Learn Real
A v1 that's in users' hands beats a perfect v2 that ships in three months. I build the smallest thing that can prove or disprove the core assumption, then iterate on signal — not on speculation.
Design Literacy Is a Force Multiplier
Being able to read a Figma file, spot a spacing inconsistency, or push back on an interaction pattern saves more time than any technical shortcut. Engineering and design working from the same mental model is where the best products come from.
“The interesting stuff is always in the tradeoffs — not in the implementation.”
Technical Arsenal
Years of experience shown on each skill.
67 skills across the full stack — frontend, backend, AI/LLM, data engineering, and eCommerce.
05 — Contact
Let'sbuild
something
remarkable.
Currently open to senior engineering roles and select freelance projects. If you're working on something interesting, I'd like to hear about it.
Say hello
Currently exploring
AI-native product tools, chess analytics at scale, and how embeddings can replace structured search in consumer apps.