Innovation & R&D
What is next, engineered into what ships.
AI / ML systems
Hypothesis: Compacted episodic memory keeps agent task success within tolerance at a lower context cost.
Applied research in AI and machine learning systems and in mixed reality, plus the products we build and run ourselves. We prototype the frontier, prove it, then transfer it into production.
- Applied AI and ML research
- Mixed reality and spatial UI
- Rapid prototyping
- Benchmarks and evaluation
- New product incubation
- Research to production transfer
01Point of view
Research that never ships is a hobby. We prototype the frontier, prove it with evidence and transfer it into production.
02Research tracks
Two tracks. One test: does it hold up?
Track A · Open questions
AI and ML systems
Agent runtimes, memory, evaluation and inference economics: the open questions that decide whether AI holds up in production.
- 01Agent runtimesHow do long-running agents stay reliable, resumable and observable when one task spans hours and dozens of tools?
- 02MemoryWhat should an agent remember, forget and cite, and how does that memory stay correct as the world changes?
- 03EvaluationHow do you prove an agent is good enough before it touches the operation, and keep proving it after?
- 04Inference economicsWhat does each decision cost, and how far can routing, caching and smaller models cut it without losing quality?
Track B · Open questions
Mixed Reality
Spatial interfaces that carry intelligence into the physical world, where people work, train and move.
- 01Spatial interfacesWhere should information live in a physical space so it helps without getting in the way?
- 02Grounded intelligenceHow does an assistant understand what the wearer is looking at, and act on it?
- 03Training and guidanceCan spatial guidance shorten the time it takes someone to learn a physical task?
- 04Hands-free interactionWhich inputs, voice, gaze or gesture, hold up in noisy, gloved and moving environments?
03Experiment pipeline
Most ideas should die cheaply.
Every experiment passes five stations, each with criteria agreed before it starts. Failing early is the point: it costs days, not budgets, and the survivors arrive in production with evidence.
01Hypothesis
A falsifiable claim about value.
- Names the decision or workflow it changes
- States what would prove it wrong
Kill signal · No measurable outcome
02Prototype
The cheapest thing that can be wrong.
- Runs on real or realistic data
- Built to learn, not to last
Kill signal · Needs heroics to work at all
03Benchmark
Measured against today's way.
- Beats the baseline on agreed metrics
- Quality, latency and cost per task logged
Kill signal · Wins only on the demo
04Pilot
Real people, real workflow, guarded.
- People choose to keep using it
- Holds up on edge cases and failure modes
Kill signal · Adoption stalls or risk rises
- ● In production
05Transfer
Hardened and shipped to production.
- Observability, evals and governance in place
- Owned by a production team
Outcome · Ships
04Products
Product 01 / 02 · Agentic project management
FinishLine
Every piece of work, carried to the finish line.
Our agentic project management operating spine. People and their agents take work from intent to delivery in one place, and we run our own company on it.
- Agents as teammates
- Claude, ChatGPT and Cursor act in FinishLine as the person who sent them.
- Two-person QA
- Nothing is done until someone who didn't build it has passed it.
- Client portal
- Clients follow progress, approve QA and raise requests.
- Two-way sync
- Slack, GitHub and Linear stay in step with the work.
FL-142Client approvals in the portal
Delivered- ✓Intent“Clients approve work in one place, not over email.”Captured from the client call
- ✓PlanningApproval flowPortal screensNotify on sign-offSprint 14 · drafted by Claude via MCP
- ✓ExecutionNKAI100%Agent acting as NK · GitHub PR #212 · Linear in sync
- ✓EvidenceQA round 1 ✕ 2 issues→Fix round→QA round 2 ✓Passed by someone who didn’t build it
- ✓DeliveryClient accepted in the portal ✓Their Slack, bridged in · release notes sent
Delivery traced back to intent ✓Gate 05 / 05
04Products
Product 02 / 02 · Agentic customer conversations
Elk.
One AI agent for all your customer conversations.
Elk learns a business from its website, docs and tools, then answers, books and qualifies across every channel, and hands over to a person when it should. Any model, retrained automatically as the content changes.
- Learns your business
- Omnichannel
- Takes action
- Model-agnostic
- Human handoff
- Web chat
- Slack
Web chatDo you ship to Dubai?
AnsweredYes, 3 to 5 days to the UAE, duties included.Cited · shipping policy
WhatsAppCan I see a demo on Thursday?
BookedThursday at 3pm is open. Booked, invite sent.Cal.com
InstagramPricing for a 20-person team?
Lead capturedHere is the Team plan. I've passed your details to sales.Google Sheets
SlackRefund for order #4471, it arrived damaged.
Handed to a personSorry about that. Looping in our support team now.Intercom ticket
Web chat¿Tienen soporte en español?
AnsweredSí, respondemos en español, a cualquier hora.Multilingual
05Incubation
Your product, incubated in our lab.
- 01 →
Frame it together
We size the opportunity with you and agree what would make it worth building.
- 02 →
Prove it cheaply
Prototypes and pilots with real users, judged against criteria set up front.
- 03 →
Build and run it
We design, engineer and operate the product with you, from launch onwards.
06Capabilities
Six disciplines. One lab.
Applied AI and ML research
Research aimed at production questions: agent runtimes, memory, retrieval, fine-tuning and model choice, tested on your data rather than on leaderboards.
Deliverables
- Research brief and findings
- Reference implementation
- Adopt, adapt or drop call
Mixed reality and spatial UI
Spatial interfaces for headsets, mobile AR and the web that put guidance, data and agents into the physical spaces where work happens.
Deliverables
- Spatial interaction prototypes
- Device and platform evaluation
- Pilot-ready MR application
Rapid prototyping
Working software in front of real users quickly, built to answer one question cheaply and honestly before anyone commits a budget.
Deliverables
- Functional prototypes
- User sessions and findings
- Go or no-go evidence
Benchmarks and evaluation
Evaluation suites that measure quality, safety, latency and cost per task against a baseline, so decisions rest on numbers, not demos.
Deliverables
- Custom eval datasets and graders
- Model and vendor comparisons
- Regression suites for production
New product incubation
Products we build and run ourselves and with partners, from framing the opportunity to a launched and operated product.
Deliverables
- Concept validation
- MVP build and launch
- Operating and growth roadmap
Research to production transfer
The hand-off where most R&D dies. We harden what worked, add observability and governance, and ship it into the real workflow.
Deliverables
- Production architecture
- Observability and eval gates
- Handover to the owning team
07Stack
Tools of the lab
Representative, not a fixed stack. We pick the instrument for the experiment: AI and ML systems on the first row, mixed reality and spatial on the second.
- PyTorch✕
- JAX✕
- Hugging Face✕
- vLLM✕
- Ollama✕
- Weights & Biases✕
- LangGraph✕
- ONNX Runtime✕
- Unity
- Unreal Engine
- WebXR
- Apple visionOS
- Meta Quest
- OpenXR
- Blender
- Three.js
08FAQ
Questions, answered.
Applied, time-boxed research aimed at a decision. We start from a question that matters to your operation, such as whether an agent can run a workflow reliably or whether a spatial interface makes training faster. Then we prototype, benchmark against criteria agreed up front and report what we found. Every experiment ends in a clear call: transfer it to production, iterate, or stop.
Transfer is designed in from the first day. Prototypes run on real data paths with evaluation suites and cost measurements, so the evidence carries over. When an experiment passes its pilot, the same team hardens it, adds observability and governance and ships it into the real workflow through our AI and application engineering practices.
Yes. Mixed reality is one of Asymmetri's two research tracks, alongside AI and machine learning systems. We design and prototype spatial interfaces that carry intelligence into the physical world, where people work, train and move, using engines and platforms such as Unity, Unreal Engine, WebXR, Apple visionOS and Meta Quest, chosen per problem.
Yes. Alongside client work, Asymmetri designs, builds and operates its own products, including FinishLine, our agentic project management operating spine, and Elk, one AI agent for all of a company's customer conversations. We bring the same model to partners: frame the opportunity together, prove it with prototypes and pilots, then build and run the product with you.
Bring us the problem, not a specification. A first conversation frames the question and what a good answer would be worth. We then propose a scoped experiment with explicit pass criteria, so you know what will be tested, how it will be measured and what happens if it works. Write to hey@asymmetri.in to begin.
AI Agentic Systems and Data↗
Systems that think, decide and act on your data.