Project
Redrob
Build India's Naver: B2C apps that compound into stronger AI
Naver is Korea's dominant consumer internet platform - search and daily services under one brand. The Redrob bet is the same shape for India: Chat, Jobs, Skills, Signal (and more) earn consented usage that makes the models and ranking better for every surface. Figures below as of the Nov 2025 Series A unless noted.
Summary
Technical co-founder and CTO since March 2018. The platform reached $6.8M ARR and 3 million users across 500 universities in India as of the Nov 2025 Series A, and later 10M+ user signups, built by an engineering organization spread across Seoul, India, and the US.
My role
Technical co-founder and CTO: owned architecture and model/cost constraints, built the engineering org from zero across Seoul, India, and the US, and still write production code. Product surfaces and regional on-call are shared with local ownership after hand-off.
Purpose
Become India's Naver: one consumer brand where careers, skills, predictions, and AI chat are daily products - not demos - and every consented interaction compounds into a stronger model and search stack for emerging markets. Free campus access ran under a fixed per-student cost ceiling and a GPU-hour budget per MAU; those two numbers were the constraint every model decision was checked against. Enterprise GTM/HR/API and cost as a hard constraint are how that flywheel stays solvent.
Timeline
2018
Company founded; joined as technical co-founder and CTO
2023
$4M seed
pre-Series A
CareerChat shipped and handed off to local ownership; later became Redrob Jobs (candidates apply; recruiters use AI for assessment)
Nov 2025
$10M Series A led by Korea Investment Partners; $6.8M ARR and 3 million users across 500 universities in India as of that release
IPL 2026
Redrob Signal launch window: free-to-play cricket prediction app; reported 2M+ downloads and ~1M sign-ups in under 60 days, with ~100K sustained DAUs - a B2C engagement surface on a parallel consented store
2026
redrob.io ships Chat, Jobs, Skills, and Signal for consumers; GTM, HR, and API for enterprise; Mail and Luma listed as coming soon
Decisions
One brand, many B2C doors into one AI
Chat, Jobs, Skills, and Signal are not separate companies. They are consent paths into the same platform: usage under consent strengthens routing, ranking, and RAG, which improves every product - the Naver-style flywheel for India.
Cost as a first-class model constraint
Free campus access ran under a fixed per-student cost ceiling and a GPU-hour budget per MAU. Those two numbers were the constraint every model decision was checked against. Quantization came first, then distillation into smaller models, then big/small routing, then the full combination. Three model tiers: a small quantized/distilled default, a mid tier, and a large tier for hard queries and paid plans. Free default stayed on the small path.
One completion surface, plan caps differ
Student and enterprise traffic share the same routed completion contract. Free and paid differ by plan caps, rate limits, and default model tier - not by a separate stack. Live routing at Series A used manual tiers and rules (plan, latency, language); the public eval harness informed offline prompts and configs rather than owning the live router.
Open Indic weights behind the contract
Open Indic weights cleared coverage, cost, latency, and license together. With open-source DNA there was little reason not to put them behind the routed contract when they passed the bar. Named upstream backends are not listed publicly.
Multilingual coverage for Indian markets
The initial quality gate was production for the languages we launched with. The goal was democratizing access for tier-2 and tier-3 cities with weak internet, not only metro English demos. Chat markets 30+ languages; Jobs supports major Indian-language search.
Text-to-ES DSL with a rubric gate
The routed LLM turns a user goal into an Elasticsearch DSL query. Generated queries are reviewed against rubrics before they hit a large professional profile index; search metrics then fine-tune those prompts offline, outside the request path. Search started on Elastic.co. When the bill became the largest line in infrastructure spend, we moved to a successor architecture and cut it by roughly an order of magnitude.
Shared workforce plane; Signal parallel
Workforce profiles, jobs, and skill scores land in Elasticsearch plus RDS/pgvector RAG on a weekly EventBridge to S3 to EMR/Glue path, with Redis for result and prompt cache. Signal prediction and ledger events sit in a parallel consented store with shared auth and consent policy. Enterprise GTM/HR/API and Console/Studio sit on the same brain; they do not fork a second one.
Organization
Hiring bar, review, and ownership
Hiring bar, review process, and ownership rules are Decisions, not only Challenges. CareerChat (later Redrob Jobs) was the first product that proved build-then-hand-off. India held primary on-call for Indic text and Jobs; Seoul escalated model and infra; US covered enterprise Console. Non-negotiable: morning standup across regions plus an agent that automates review.
Challenges
Consent is the product, not a footer
A Naver-scale flywheel only works if users understand what they give and what they get. Predictions, resumes, and searches stay permissioned; trust failure kills the data advantage.
Free access without unbounded unit cost
Abuse and moderation broke first under free load, which cascaded into GPU failures and then support tickets on degraded quality. First mitigations were rate limits and stronger moderation gates, then a smaller free-tier default so remaining GPU capacity served real users. Structural distillation and edge work followed after the fire was out.
Quality that survives real Indian use
Indic tokenizer fertility was a production cost problem: demo spend looked manageable, then production spend blew up by roughly the same order of magnitude as the fertility gap. Tokenizer blow-up, script mixing, and weak local-entity handling killed candidate stacks. English-benchmark-only ships were blocked until they cleared the Indic production gate and India product tasks (Jobs/Chat). Enterprise and student used two quality bars: enterprise needed P99.9 stability and low latency under SLA; student optimized for reach inside the per-student cost ceiling and GPU-hour budget.
One brand, many daylight windows
Timezone handoffs and review bottlenecks cost the most calendar time across Seoul, India, and the US. Build-then-hand-off rarely reversed; when it did, Seoul took infra or model-weight ownership while India kept Jobs/CareerChat UX, usually forced by GPU or cost incidents. Morning standup and the review agent exist so ownership does not drift.
Impact
Revenue and reach
$6.8M ARR (Nov 2025); 10M+ user signups across 500 universities in India (3 million users as of the Nov 2025 Series A).
Redrob Signal (IPL 2026 launch)
Free-to-play skill-based prediction app (cricket first; entertainment and economy topics expanding). Launch-window reports: 2M+ downloads and ~1M sign-ups in under 60 days, ~100K sustained DAUs for a month, #1 Free Sports on Google Play in India, with AB de Villiers and Dale Steyn as public endorsements. Not real-money gaming - a high-engagement B2C surface on a parallel consented plane.
Funding
$14M raised ($4M seed, $10M Series A). As of the Nov 2025 Series A.
Organization
An engineering organization built and trained across Seoul, India, and the US (0 to 50+) inside a 100-person company operating from five cities (San Francisco, New York, New Delhi, Mumbai, Seoul) as of the Nov 2025 Series A.
Public product suite
redrob.io publishes consumer Chat, Jobs, Skills, and Signal; enterprise GTM, HR, and API; campus access; Mail and Luma as coming soon. Together they are the Naver-shaped surface area - one brand, many daily reasons to return.
System
Redrob platform
3 tiers
routed · cost ceiling + GPU-h/MAU
India's Naver-shaped flywheel: B2C Chat/Jobs/Skills/campuses and enterprise GTM/HR/API share one routed completion contract (plan caps differ). Workforce data in Elasticsearch (large professional profile index; Elastic.co era, successor undisclosed) + RDS/pgvector RAG; Signal on a parallel consented store. Redis caches results/prompts. Live routing: manual tiers/rules (not GEPA live). Weekly EventBridge→S3→EMR/Glue ingest. Offline rubric metrics fine-tune prompts. AWS: CloudFront/WAF/Shield Standard, ECS+EKS, private VPC. Three model tiers (small quantized/distilled default, mid, large for hard/paid); free default on small path under a fixed per-student cost ceiling and GPU-hour budget per MAU.
B2C apps and parallel Signal store under consent → AI core (routed LLM with plan/rules tiers, rubric gate, Elasticsearch, pgvector, Redis) → stronger models back into every surface; weekly EventBridge/S3/EMR ingest; Shield Standard at the edge
Screens captured from public redrob.io marketing and product surfaces (Jul 2026). People Search is an in-product UI; Signal is the consumer prediction app page; other shots are product pages that embed live UI previews. Console and Studio have their own project pages.
Screens
Brand homepage
redrob.io positions one brand over a product suite for careers, communication, learning, and everyday decisions. Primary free starts are Chat, Jobs, and Skills.
Redrob AI Chat
Flagship LLM product page: India's AI for the next billion professionals, built on a large professional profile index and 30+ languages for hiring, sales, jobs, and research in one system.
Conversation surface
Public product framing: multilingual coverage, a small-model path, and one conversation that reaches people search, company search, job search, and resume ranking.
People Search
In-product people search over a large professional profile index with filters and contact unlock. This is the workforce-data surface behind Chat and GTM/HR workflows.
Redrob Jobs
Resume-first job search: upload once, get ATS score and ranked listings from LinkedIn, Naukri, Wellfound, and other sources, with an AI agent path to recruiters.
Redrob Skills
Verified skills tests with a Redrob Score out of 1000, portable credentials, and recruiter/institution modes. Completes the student-to-hire loop with Chat and Jobs.
Redrob Signal
Free-to-play prediction app: predict match outcomes, earn points, climb leaderboards. Cricket-first during IPL; entertainment and economy topics expanding. Not real-money gaming.