I build web products, automation systems and ML tools that survive contact with production — from React front-ends and Node services to RAG pipelines and the occasional toy compiler.
I treat every product like a survival world — scout the terrain, craft the right tools, and build things that are still standing after the first night in production.
Four years of building took me across the whole map: React and Next.js front-ends, Node services, Zoho automation farms for real sales teams, and ML models that had to run on cheap phones — not just in notebooks. I like the unglamorous parts: error states, migrations, edge cases. That's where products actually live or die.
Lately I've been tunneling downward — writing a toy language in Rust and reading the Go runtime for fun. Going lower made the upper layers make more sense, not less.
24+ builds shipped · 200+ CRM automations live · 3 RAG systems in production · 1 toy compiler (and counting)
The tools I actually reach for, sorted by category. Proficiency is self-graded and honest — nothing here is "familiar" in the résumé sense of the word.
HOVER A SLOT TO INSPECT THE ITEM
ENCHANTS REFLECT PROFICIENCY
Five projects that survived contact with users — hover to preview, click to open the crafting notes.
A retrieval layer that indexes a team's scattered docs, wikis and tickets, then answers questions with citations instead of links. Local sentence-transformer embeddings, pgvector storage, and a Next.js front-end that streams responses token-by-token. Built for the "where is X documented?" problem — the one that quietly eats engineering hours.
OUTCOME — internal "where is…" messages dropped ~70% for a 15-person team.
The full automation layer for a growing sales org: 200+ Deluge workflows, lead-scoring functions, cross-module blueprints and nightly reconciliation — plus a Node sidecar service for everything Deluge couldn't do cleanly. Failure alerting and audit trails from day one, because silent automation is a liability, not a feature.
OUTCOME — sales ops cut manual entry ~60%. Zero silent failures since alerting shipped.
An expression-first toy language with a lexer, Pratt parser and bytecode VM, written in Rust and compiled to WASM so the playground runs fully client-side. Error messages were treated as a product feature, not an afterthought — every parse failure points at the exact character and suggests a fix.
OUTCOME — ~900 LOC of core · 120+ passing tests · playground works offline.
A TFLite image classifier fine-tuned on plant-pathology datasets, wrapped in a lightweight app built for low-connectivity districts. Int8-quantised down to 11 MB, inference in ~180 ms on budget Android hardware, with a full offline cache — the field is the only network it needs.
OUTCOME — 87% top-1 across 12 diseases on held-out regional validation.
Seeded value-noise terrain, greedy-meshed chunks, fog and a small day-night cycle — a weekend flight into engine work that turned into a month. The world is deterministic from its seed, effectively infinite in X/Z, and generated on the fly as you walk. (Yes, the floating island in the hero is a cousin.)
OUTCOME — holds 60 fps with ~250k visible blocks on a mid-tier laptop.
Every role was a biome with its own hazards. The full log, most recent first — the torch follows your scroll.
Shipping end-to-end for founders and small teams: Next.js products, Node/Postgres backends, and RAG pipelines with real evaluation loops. Also the person they call when a Zoho automation farm has quietly rotted for two years.
Designed the automation layer for a 40-seat sales org: 200+ Deluge workflows, lead-scoring functions, nightly reconciliation and failure alerting, with a Node sidecar for everything Deluge couldn't do cleanly. Ops overhead dropped by roughly 60%.
Built and maintained React/Next front-ends and Node APIs across eight client launches. Learned that most "ML features" die in deployment, not in modeling — and started fixing that side of the pipeline.
First boss fight. Shipped internal tooling, a TFLite prototype for document triage, and a durable habit of writing down every production mistake.
Officially: algorithms, DBMS, operating systems. Unofficially: compilers, graphics, and far too many side projects. Left with a degree and a much longer reading list.
The portal's open — for freelance work, full-time roles, or a good argument about compilers. I read everything and usually reply within a day, IST.
STATUS — OPEN · FREELANCE / FULL-TIME · IST (UTC+5:30)
CLICK TO COPY · NO FORMS, NO FUNNEL
ELSEWHERE
> aasim.dev — v2.4.1
> generating terrain .......... ok
> spawning player ............. ok
> loading chunks [12, 84] ..... ok
seed 1189787 · render distance 12 chunks