We build AI products, from the model to the app.

Agents, voice, retrieval and custom models, in apps people use. AI drafts the code, our engineers decide what ships.

How a change ships at Dravin AIBugs and feature requests arrive on one intake board. Five AI coding agents work in parallel, each in its own isolated environment and branch, and tests, security scans and a preview deployment run on every pull request. The changes then meet one of our engineers, who reads each change, sends work back and approves the pull request. Approved changes merge into main under your branch protection, then build, test, deploy to staging and deploy to production automatically.How a change ships at Dravin AIBugs and feature requests arrive on one intake board. Five AI coding agents work in parallel, each in its own isolated environment and branch, and tests, security scans and a preview deployment run on every pull request. The changes then meet one of our engineers, who reads each change, sends work back and approves the pull request. Approved changes merge into main under your branch protection, then build, test, deploy to staging and deploy to production automatically.
Book a callSee how we work

30+ production AI systems shipped

Some of the companies we’ve worked with

What we’ve built

From an agent in your inbox to your own model on your own GPUs.

01

Agents and automation

Agents that plan, call your tools, remember and run on a schedule.

  • Tool-calling agents, MCP servers and A2A handoffs
  • Agents for sales, support, finance, HR and ops
  • OpenClaw in WhatsApp, Slack or Teams, locked down
  • One chat window driving 100+ agents
  • MCP
  • A2A
  • OPENCLAW
  • LONG-TERM MEMORY
  • CLAUDE CODE
Agents in depth
02

Content and generative media

Content that ships every day in your voice, in Hindi, regional languages and English.

  • SEO and AI-search content, published to your CMS
  • Social, newsletters and ad variants, approved before posting
  • Product images, AI video, avatars and dubbing
  • Hindi, regional languages and English, one brand voice
  • VEO 3.1
  • KLING 3.0
  • HEYGEN
  • FLUX.2
  • NANO BANANA 2
  • ELEVEN V3
Content in depth
03

Voice agents

Voice agents that answer calls, start speaking before the model has finished and stop when the caller cuts in.

  • Support lines, bookings, sales calls and pre-screens
  • Turn detection, so callers can cut in
  • Indian and European languages, self-hosted if you need
  • Whisper in, Kokoro out, served on Triton
  • WHISPER
  • KOKORO
  • TRITON
  • GPT-REALTIME-2
  • DEEPGRAM FLUX
  • LIVEKIT AGENTS
Voice agents in depth
04

Your own AI

Models trained on your data, for your jobs, owned by you.

  • Continued pretraining, SFT and RL post-training
  • Distilled into small, fast models that hold the score
  • From a phone-sized model to open MoE giants
  • A 3B model that beat a frontier API model on the client’s own benchmark
  • PYTORCH
  • TORCHTITAN
  • VERL
  • TRL
  • DEEPSEEK V4
  • QWEN3.8
Training in depth
05

Inference and on-device

Models served where your data lives, at a lower cost per token.

  • On the phone, the laptop or in the browser, offline
  • Private GPUs and air-gapped networks
  • Routing, caching, batching and quantisation
  • Large MoE models with disaggregated serving
  • VLLM
  • SGLANG
  • TENSORRT-LLM
  • MLX
  • EXECUTORCH
  • CORE AI
Inference in depth
06

Beyond LLMs

A small trained model for labels, forecasts and images; an LLM where language is the job.

  • Classifiers for routing, triage and guardrails
  • Forecasts for demand, pricing and inventory
  • Vision: detection, inspection and document forensics
  • Demand forecasting on 150+ engineered features
  • XGBOOST
  • LIGHTGBM
  • TABPFN
  • CHRONOS-2
  • SAM 3
  • SHAP
Beyond LLMs in depth
07

Apps and business software

Apps and the software your business runs on, with AI built in.

  • Web, iOS, Android, Mac and Windows apps
  • CRMs, ERPs, inventory and self-hosted project trackers
  • Learning platforms, portals and internal tools
  • A mobile AI app with 30k users
  • NEXT.JS
  • REACT NATIVE
  • EXPO
  • SWIFTUI
  • ELECTRON
  • FASTAPI
Apps in depth
08

Fixing AI-built apps

AI-built apps with real users, made safe to grow, usually without a rewrite.

  • Secrets moved out of the repository
  • Tests on the paths that make money
  • Deploys you can roll back
  • Checked by our own autonomous web-testing platform
  • LOVABLE
  • BOLT
  • CURSOR
  • REPLIT
  • V0
  • CLAUDE CODE
Cleanup in depth
See every capability

Screens from software we have shipped.

Screens rebuilt with sample data. No client data is shown.

Medical coding engine

Documents and knowledge

A coder’s review of one outpatient visit: the clinical note with evidence highlighted, suggested diagnosis and procedure codes, the rules engine’s verdicts and the visit level.

A medical coding engine that assigns diagnosis, procedure and visit-level codes from clinical notes, with a rules engine making the final call.

  • HYBRID RETRIEVAL
  • RULES ENGINE

SME credit-underwriting platform

Agents and automation

The credit appraisal for one SME loan: the three-year financial spread against policy, the platform’s risk signals, bureau and banking, and the recommended limit with its conditions.

An SME credit-underwriting platform: KYC and director checks, litigation and regulator screening, adverse-news analysis, financial ratios and a cited document chat, all exposed as MCP tools.

  • LANGGRAPH
  • MCP
  • PADDLEOCR
  • QDRANT

Edtech platform

Apps and business software

A student lesson on velocity-time graphs: an interactive simulation, a quiz question and a question asked in Hindi, answered on the graph.

An edtech platform launching to 50k students: lesson planning, gradebooks, attendance, a tutor for students’ questions and a studio that generates interactive simulations.

  • WEB PLATFORM
  • BULLMQ
  • OPENTELEMETRY
What we’ve built

What gets caught before a change merges.

Automated checks run on each pull request, then an engineer reads the change.

  1. N+1 queriesA database call per row becomes one batched call.
  2. InjectionUser input built into SQL becomes a parameterised query.
  3. Missing types or testsLoose types get real ones, and the paths that matter get tests.
src/orders/get-user-orders.ts7 lines removed, 11 lines added.
  • sql injection
  • n+1 query
async function getUserOrders(
ids: string[],
) {
removed: for (const id of ids) {
removed: const user = await db.query(
removed: 'SELECT * FROM users ' +
removed: 'WHERE id = ' + id);
removed: const orders = await db.query(
removed: 'SELECT * FROM orders ' +
removed: 'WHERE user_id = ' + id);
added: const [users, orders] =
added: await Promise.all([
added: db.query<User>(`
added: SELECT * FROM users
added: WHERE id = ANY(:ids)`,
added: { ids }),
added: db.query<Order>(`
added: SELECT * FROM orders
added: WHERE user_id = ANY(:ids)`,
added: { ids }),
added: ]);
See the full process

From the data to the deploy.

What we build with: 64 tools in 8 groups. The full list follows.

Classical ML and data

  • scikit-learn
  • pandas
  • NumPy
  • Polars
  • Jupyter
  • Optuna
  • MLflow
  • XGBoost
  • LightGBM
  • CatBoost
  • AutoGluon
  • SHAP

Training and serving

  • PyTorch
  • Hugging Face
  • NVIDIA
  • vLLM
  • Triton
  • Ollama

Models

  • OpenAI
  • Claude
  • Gemini
  • Meta
  • Mistral AI
  • DeepSeek
  • Qwen

Agents and memory

  • LangChain
  • MCP
  • PostgreSQL
  • Qdrant
  • Neo4j
  • Redis
  • MongoDB

Voice and media

  • ElevenLabs
  • Deepgram
  • LiveKit
  • Twilio
  • Whisper
  • Kokoro TTS

Product and apps

  • React
  • Next.js
  • Vite
  • Node.js
  • FastAPI
  • Fastify
  • Supabase
  • Tailwind CSS
  • Expo
  • Electron

Languages

  • Python
  • TypeScript
  • JavaScript
  • Go
  • Rust
  • Swift

Cloud and DevOps

  • AWS
  • Google Cloud
  • Azure
  • Cloudflare
  • Vercel
  • Docker
  • Kubernetes
  • Terraform
  • GitHub Actions
  • Playwright
  • 100k+ customers on an agent platform we built for
  • 30k users on a mobile AI app we built
  • An edtech platform we built, launching to 50k students

What we have been writing about.

Read the blog

Tell us what you want to build.

A call is the quickest way to start. If you already have code, request a code review and get a written report on it within 48 hours. Read a full report.

Team
Core team from IIT, led by founders who are researchers turned engineers from IIT Kharagpur.
Experience
Two years serving customers, from startups to enterprise teams.
Shipped
30+ production AI systems.
Book a call