Applied AI · TypeScript · EU-based
Production-grade AI features
for Series A startups.
Built end-to-end in TypeScript.
Mythal Studio ships AI features that move the business — RAG, semantic search, embeddings, rerankers. Production code your team can own and extend.
Applied AI, shipped end-to-end.
01 — Services
Services
Three ways to engage, ordered by commitment. Pricing on request — every brief comes back with a fixed quote.
- 01
AI Feature Audit
1–2 weeksA focused review of an AI feature already in your roadmap or in production. Maps risks, identifies the cheapest path to working, and returns a written plan with concrete tradeoffs.
Written audit · Architecture sketch · Cost estimate
Pricing on request
- 02
Integration Sprint
4–8 weeksEmbeds with your team to ship one critical AI feature into production. RAG, semantic search, embeddings, rerankers, or an LLM-driven workflow. PRs against your repo, code your team owns afterwards.
Production feature · Evals · Runbook
Pricing on request
- 03
Embedded Engineer
3+ monthsSenior engineer plugged into the product roadmap for a quarter or more. AI features alongside the rest of the work. Async, on CET hours, with one fixed weekly checkpoint.
Senior bandwidth · Roadmap input · Continuity
Pricing on request
02 — What gets built
Every layer of the applied AI stack
Twelve concrete things Mythal Studio builds end-to-end. Each one ships to production with evals and a runbook your team can extend.
- 01
Semantic search & RAG
Search and assistants over your knowledge base, with answers grounded in citations. Hybrid retrieval, reranking, ingestion pipelines, and evaluation built in.
- 02
Product copilots
Context-aware assistants embedded in your product UI. Streaming UX, tool use against your APIs, and guardrails so the copilot stays inside the rails.
- 03
Multi-step agents
Agents that plan, call tools, validate, and hand off to humans when needed. Built on LangGraph or hand-rolled, chosen by what fits the workload.
- 04
Voice agents
Telephony or in-product voice agents that qualify, schedule, or resolve. Whisper, TTS, Twilio, and streaming pipelines wired to your data.
- 05
Recommendations & personalization
Hybrid vector + behavioral recommenders for product, content, or feed. Offline eval harness and A/B flag baked in from day one.
- 06
Document understanding
Structured data out of PDFs, scans, and forms via vision + LLM. Schema-validated outputs, batch pipelines, and accuracy tracked over time.
- 07
Workflow automation
Agents that triage, draft, classify, and route the work humans repeat all day. Native connectors to Slack, Jira, Linear, ServiceNow, and the rest.
- 08
Observability & evals
Tracing, cost tracking, and regression eval suites for production LLM features. Quality stops being a vibe and becomes a number on a dashboard.
- 09
Backend & APIs
Production services around your agents: auth, rate limiting, queues, webhooks, streaming. TypeScript end-to-end, Postgres-grounded, Vercel or AWS.
- 10
Greenfield products
AI-first products from zero: landing, auth, billing, dashboard, the AI feature, observability. One team owns the whole thing through to launch.
- 11
Data pipelines for AI
Ingestion, normalization, embeddings refresh, evaluation datasets — across Postgres, S3, Notion, Drive, and the messy systems no one wants to touch.
- 12
Modernization & acceleration
Refactor legacy platforms and add AI without destabilizing production. Incremental migration, parallel runs, clear rollback at every step.
One studio. One point of accountability. No handoffs to a platform team that doesn't exist yet.
03 — Case studies
Case studies
Selected engagements, anonymized to honor NDA terms. Metrics measured against pre-deployment baselines.
- 01Under NDA
Investor research platform · Series A
Hybrid RAG over 12k+ expert interviews
Semantic + lexical retrieval pipeline for an investor research platform, with long-form transcript chunking, a cross-encoder reranker on the top-200 candidates, and citation-grounded answers wired into the analyst workflow. Postgres + pgvector, OpenAI embeddings, Anthropic Claude.
-78% time to first relevant insight, 3.2× recall on previously-missed transcripts.
- 02Under NDA
Life-event vendor marketplace · scale-up
Personalized vendor ranking for a marketplace
Three-layer personalization pipeline for a scale-up marketplace, with candidate generation over vendor embeddings, a learned ranker balancing budget, region and style fit, and a real-time session model picking up intent shifts within the same browse. Postgres + pgvector, ClickHouse, learning-to-rank.
+11 pp session conversion in A/B, +14% revenue per visitor on the test cohort.
- 03Under NDA
Regulated sports-betting operator
Real-time collusion detection inside compliance
Account-linkage features fused with embedding similarity over historical risk patterns, scored by a lightweight classifier and slotted inside the existing compliance flow with audit-grade citations. Postgres + pgvector, Python classifier, Helicone.
-68% manual review time, +22 pp recall on previously-missed collusion patterns.
04 — How we work
How we work
A four-phase delivery model. Senior-led, evidence-driven, built for production from day one.
- Phase 01
Discovery
Goals, systems, and delivery risks mapped before the scope hardens.
- Phase 02
Solution design
Data, security, success metrics, and pilot scope, defined with clear acceptance criteria.
- Phase 03
Validation & build
Value is proven in a pilot first, then scaled into a full build with agent-assisted delivery and senior review.
- Phase 04
Deployment & support
Production rollout with runbooks, monitoring, and outcome reviews tied to the agreed KPIs.
05 — Stack
Stack
The tools, frameworks, and platforms shipped on every day. If your team already uses them, the studio already knows them.
Agents & frameworks
- LangGraph
- CrewAI
- LangChain
- Tool / Function calling
- MCP (Model Context Protocol)
- A2A Protocol
- Planner-Executor
- ReAct
- Chain-of-Thought / Tree-of-Thought
- Prompt engineering
LLM platforms
- OpenAI
- Anthropic
- Google Gemini
- AWS Bedrock
Memory & retrieval
- Postgres + pgvector
- Qdrant
- Redis (vector)
- Hybrid search
- RAG
- Rerankers
Backend & runtime
- TypeScript
- Next.js App Router
- Node.js
- Bun
- Deno
- Hono
- tRPC
- Drizzle ORM
- Python
- FastAPI
Infra & deployment
- AWS ECS / Fargate
- Vercel
- Cloudflare
- Edge runtime
- Docker
- GitHub Actions
- Kafka
- Celery
Voice & multimodal
- Whisper / TTS
- Twilio
- Multimodal pipelines
- Streaming UI
- WebSockets
Integrations
- Slack / Teams
- Jira
- Linear
- ServiceNow
- SharePoint
Observability & evals
- Helicone
- Tracing
- CI/CD
Security & compliance
- Prompt injection defense
- PII redaction
- RAG poisoning prevention
- OWASP LLM Top 10
- EU AI Act
- GDPR
06 — About
About Mythal Studio
Mythal Studio is a senior-led practice for applied AI engineering, working with Series A startups across the EU. Built and operated from Madrid, async-first, on CET hours. The focus is integrating AI where it moves the business: features users feel, owned by your team.
07 — Send a brief
Tell Mythal Studio what to ship.
A short note about what you're building and where the team needs help. A written reply with next steps, a rough approach, and how soon work could start — usually within two business days.