Systems & Deep Learning AI Engineer
Not specialized in a stack — specialized in a repeatable process across architecture, cloud infrastructure, and applied deep learning. Built for real users, with the failure modes on the record.
Selected work
Nine systems, nine different problems, one process.
Each row expands into the spec: what it does, the architecture behind it, the stack, and — deliberately included — what doesn't work yet. Tap a row.
Data flow
What it does
Reads contact-sport footage through pose estimation and temporal modeling, then reasons over it with a four-agent mesh: Raider, Defense, Simulation, and Referee agents — addressing the occlusion problem that breaks single-model approaches.
Result
94.2%Confidence at 38ms inference, research-stage. Not yet verified at production scale.
Stack
- Python
- PyTorch
- HigherHRNet
- ST-CGNet
- U2Diff
Known limitation
Accuracy drops on low-light, low-bitrate amateur footage. Augmented-dataset training is planned to close the gap.
Data flow
What it does
Desktop ERP covering 20L-canister inventory, invoice generation, purchase/sales flows, maintenance logs, dashboards, and reports — built to run fully offline on-site.
Result
100%Local uptime, 0 cloud dependency. Solo-built and delivered in 6 weeks, including on-site deployment and end-user training.
Stack
- Tauri v2
- Rust
- SQLite
- React
- TypeScript
- Next.js
Known limitation
Simultaneous offline edits need manual conflict resolution today. A vector-clock resolution model is in development.
Data flow
What it does
Lets citizens file and track complaints and get municipal updates. A Gemini-backed chatbot triages incoming issues and routes them correctly before a human sees them.
Architecture
4 decoupled services on a RabbitMQ event bus, orchestrated with Kubernetes — no service calls another directly.
Stack
- Go
- Node.js
- Python
- Flutter
- Vue.js
- Docker
- Kubernetes
- RabbitMQ
Known limitation
High-volume complaint bursts still cause temporary queue backlogs. Auto-scaling is being added to absorb them.
Data flow
What it does
API testing built for engineers who want collections that diff cleanly in git — markdown schemas instead of a proprietary blob, plus a headless runner for CI.
Positioning
Resource-efficient and terminal-first, built as the tool its author wanted and couldn't find.
Stack
- Go
- Wails
- React
- TypeScript
Known limitation
Unsigned macOS binary trips Gatekeeper. A documented xattr workaround ships with the release notes.
Data flow
What it does
Unifies multiple LLM vendors under one /v1/chat/completions endpoint, round-robin routing across providers so a single vendor's rate limit never stalls a request.
Why it exists
Every serious multi-model project rebuilds this exact proxy once. Valluvan is that proxy, done once, self-hosted.
Stack
- Go
- SQLite
- HTMX
Known limitation
SQLite state is wiped on reboot in ephemeral cloud environments — needs a persistent volume for durable deploys.
Also shipped
Thirueadu
Offline-first personal finance ledger, double-entry accounting, local-first UX.
- Tauri v2·
- Rust·
- SQLite
Sunsea CRM
Recruitment CRM modeled on real day-to-day agency operations.
- Node.js·
- PostgreSQL·
- React
VyExpense
Trip and expense approvals with receipts, audit trails, and finance dashboards.
- Express·
- Vue 3·
- Chart.js
YogaTube
YouTube-style browsing client — a study in responsive layout and predictable state.
- React 19·
- Vite·
- Tailwind CSS
How the systems are put together
One shape, reused on purpose.
CivicConnect's event-mesh: five services, zero direct calls.
Decoupled services talk through an event bus, never directly. CivicConnect runs four services this way over RabbitMQ — a queue backlog degrades gracefully instead of one service taking the rest down with it.
Complex reasoning gets split across specialized agents instead of one model doing everything. K-RAID uses four — Raider, Defense, Simulation, Referee — each owning one part of reading the footage.
Local SQLite plus IPC keeps the app usable with no connection at all. The Water ERP runs at 100% local uptime this way — the tradeoff is manual conflict resolution on simultaneous offline edits, still being closed.
Permissions and status transitions are schema decisions from day one, not middleware added after launch — the difference between an audit trail and a guess.
How I work
Three rules I don't break,
and the process behind them.
Workflow comes first. Interface is what's left over.
The data model and the process it encodes get designed before a single screen does. UI decisions follow from that, not the other way round.
Access control isn't added later. It's part of the model.
Permissions and status transitions are schema decisions from day one — retrofitting RBAC onto a live system is how audit trails get holes in them.
If code and docs disagree, the docs are wrong. Fix immediately.
Documentation lives with the code it describes. A stale doc is treated as a bug, not a chore for later.
Operating sequence, every project
Experience & internships
Where this was tested for real.
- Designed a 17-table schema for workfast.ai's Leave & Attendance system — ledgered leave balances, effective-dated working-day rules, recurring holidays — integrated into live user, company, and check-in tables without breaking existing flows.
- Shipped 35 endpoints covering multi-level approvals resolved from the org chart and timezone-aware attendance classification, kept consistent by cron jobs for incremental sync, nightly reconciliation, and year-end carry-forward.
- Added voice-note and attachment support to a production Check-in Task Summary endpoint (Go, Gin) — moved it to multipart uploads with a JSON fallback, running CDN transfers concurrently across goroutines to keep response times flat.
- Built a PHP portal for managing online courses and external certifications, with role-based dashboards for administrators and coordinators.
- Automated the data backup process with cron jobs and shell scripts, reducing manual intervention and improving reliability.
Technical range
Not a stack.
A process,
applied across stacks.
Grouped by what each does in a system, not ranked by preference.
Languages
- Python
- Go
- Rust
- TypeScript
- JavaScript
- SQL
- Dart
- PHP
Client / interface
- React 19
- Next.js
- Vue.js
- Flutter
- Tauri v2
- Tailwind
Backend / systems
- Node.js
- Express
- gRPC
- Wails
Data & messaging
- PostgreSQL
- SQLite
- Redis
- RabbitMQ
Infra & AI/ML
- Docker
- Kubernetes
- NGINX
- PyTorch
- Gemini
- HigherHRNet
Get in touch
Read every email.
Open to contract and full-time work in systems engineering, applied deep learning, cloud architecture, offline-first ERPs, and event-mesh design.
ryogasimman@gmail.com →