An engineering team for ambitious AI-era companies. We've shipped products to 30+ countries and run engagements for fintech, agencies, and consultancies in the UK, EU, and India.
Domains we've shipped in
Client names withheld under NDA · references on request
Citation-enforced retrieval, multi-model routing, and continuous evals on Langfuse — so the model stops being a guess.
Kafka Streams, Postgres + pgvector, Cloudflare Workers, AWS SST. Architectures that survive the IPL match-day spike, not just the demo.
Three.js explainers, voice-first UIs in Hindi/English/Hinglish, Next.js 15 with strict TypeScript and design that doesn't look generated.
Prototype to production
We don't hand off a half-finished prototype. Every engagement ends in production code, observability, and a runbook your team can operate. Founder-tier engineers from day one — no juniors hidden behind the brand.
Two-week paid audit ending in a written diagnosis with a sequenced fix list. You own the document.
Two-week sprints with a Friday demo. Production-grade tests and observability shipped from sprint one.
We can run it for you, or hand the keys to your team with full documentation when ready.
Boring tech where boring works, new tech where new genuinely wins. We don't bill for slide decks.
Engineering depth
Our pattern: cite, verify, refuse. Every claim ships with a chunk-ID. Answers without retrieval support are rejected at the streaming boundary, not surfaced to the user.
# RAG with citation enforcement
async def answer(query: str) -> Answer:
chunks = await retrieve(query, k=12, rerank="cohere-v3")
if len(chunks) < 3:
return Answer.refuse("insufficient grounding")
model = router.choose(query, budget="basic") # claude / gpt-4o / gemini
stream = await model.stream(prompt(query, chunks))
async for token in stream:
if citation_violated(token, chunks):
return Answer.abort("hallucination detected")
yield tokenBuilt & running
Sunja's own products — the proof that we build and operate AI-native software at scale, not just consult on it.
Citation-enforced RAG over 10K+ exam questions. Multi-model routing across Claude / GPT-4 / Gemini cuts AI cost 67%. Voice-first TutorPortal in Hindi, English, and Hinglish.
K-12 textbook chapters as cinematic Three.js worlds — electromagnetism, mitosis, planetary motion — with an embedded AI tutor that grades open-ended answers and resolves doubts mid-lesson.
Six-stage LLM pipeline ingests, deduplicates, and editorially grades news. Three-engine adaptive crawler (Static → Playwright → Stealth) with per-domain learned profiles. Confidence-scored auto-publish gate.
Capabilities
Every line below is something we've shipped to production users — not a list copied from a job board.
Citation-enforced RAG. Multi-model routing across Claude, GPT-4, and Gemini. Semantic caching over Postgres + pgvector. Continuous evals on Langfuse that gate every prompt change in CI.
How we engage
We're a small senior team. Pick the shape that fits the work. Each option has a clear scope, deliverables, and exit criteria — no hourly billing surprises, no slide decks billed for.
Talk to us2–4 senior engineers, full ownership of a product surface, daily standups with your team.
Clearly bounded deliverable — RAG pipeline, mobile app, infra rebuild — for a fixed fee and timeline.
Founder-level architecture, hiring loops, and unblocking — without the full-time hire.
A paid diagnostic with a sequenced fix list. You own the document either way.
Selected work
We don't ship demos. Every project below went to production with paying users, regulated workloads, or both. Client names withheld under NDA.
An SMB cash-flow forecasting platform — bank aggregation, transaction analytics, runway calculation, variance analysis. We led the architecture through three iterations: JavaScript Express, TypeScript port, then a polyglot v2 splitting hot paths into SST-on-AWS serverless functions and a Python analytics engine triggered by SQS after every bank sync.
UK Open Banking via Moneyhub. MySQL schema entirely under Liquibase with branch-environment alignment. LLM transaction-name cleaning evaluated on Langfuse. Five test suites with a "critical" pre-commit gate.
Migrated a Sheets-backed founders-network portal to a hardened Supabase stack hosted in EU Frankfurt. 19 SQL migrations covering OWASP hardening, performance indexes, audit sessions; PII gated behind a server-side Edge Function.
Salesforce ↔ JobLogic ↔ workforce-locations integration on Hono + Deno Edge Functions. Content-hash idempotency, source-destination loop detection, partial-unique indexes for one-primary-per-case.
A self-hosted GrowthBook alternative on four Cloudflare Workers across 300+ edge POPs. Three-tier cache (edge / KV / R2) keeps p95 under 50ms. Capacity-planned for IPL match-day push campaigns delivering 8–10K peak RPS.
Real-time profit ops for multi-channel sellers. 44+ Cloud Functions feeding ETL into BigQuery, Terraform IaC, Cloud-Armor DDoS protection. Profit dashboards that update in seconds.
Capture-the-flag assessment for junior full-stack engineers — IoT sensor dashboard with planted bugs, a stubbed AI route, 5 flags. Comes with rubric and confidential evaluator guide. Replaced their generic LeetCode loop entirely.
What you get
Most of what's billed in this market is junior labour and slide decks. We charge for senior engineers shipping production code. No padding, no bench, no farmed-out subcontractors.
Founders
BIT Mesra computer scientists with a combined 15+ years across IoT, mobile, ML, real-time data, and AI platforms.
Mobile and IoT engineer with eight years of taking systems from prototype to production scale — most recently as AI Engineer at TVS Motor, earlier at TagBox (acquired by TVS, National Startup Award 2021). At Sunja, leads product across iOS and Android — from concept and brand through native engineering.
Platform engineer with seven years of cloud-native and AI work across TVS Motor and TagBox (acquired by TVS, National Startup Award 2021). At Sunja, leads engineering on Deep Mentor's RAG, multi-model AI routing, and the platforms behind it. Also serves as fractional CTO for an e-commerce optimization consultancy.
We reply to every inbound within one business day. If we're not the right fit, we'll point you to a team that is.
