LLM & Agentic AI
AI that survives contact with real users.
Aliencode Tech builds LLM and agentic AI systems that hold up in production: retrieval-augmented generation grounded in your own documents, agents that take real actions against real systems, and the guardrails and evaluation that keep their answers inside the boundaries you set.
Key facts
| We build | RAG systems, sales and customer support assistants, agents for education, data pipelines, automated testing, document and media pipelines and more |
|---|---|
| Grounding | Embedding-based semantic search over your own corpus |
| Surfaces | Web app, Google Chat, Slack, WhatsApp, Telegram, or your own API |
| Infrastructure | Google Cloud, AWS, or on-premise on hardware we can also supply |
| Pricing | By quotation, after scoping |
What we build
What kinds of AI systems do we build?
Retrieval-augmented generation over your own documents, sales and customer support assistants, agents for education, agents that act against your systems, the data infrastructure that feeds them, automated testing, media understanding pipelines, and the evaluation harnesses that keep all of them honest as they change.
And fully bespoke systems that fit none of those descriptions. Vigil answers one line: read every video posted on a topic, in several languages, and tell an analyst what changed since yesterday. Nothing off the shelf did that, so we built it up from the media and collection pipelines we already ran. That is usually where the interesting work starts.
Retrieval-augmented generation
Your documents become the answer source. Embedding-based semantic search finds the relevant passages, and the model is constrained to answer from them rather than from whatever it happens to remember.
Sales and customer support assistants
Assistants that answer from your own product, pricing and policy material, hand over to a person at the point they should, and log every question they could not answer so the gaps get closed.
Agents for education
Tutors and study assistants grounded in a school’s own curriculum and notes, answering in the register the syllabus is marked against and generating structured or visual output when prose is not what the student needs.
Agentic workflows
Systems that do not just answer but act (calling your APIs, filling records, triggering jobs), with the tool boundaries, retries and audit trail that make that safe to switch on.
Data pipelines at scale
Collection from your own systems, partner feeds and public sources, with the scheduling, deduplication and storage that turn a stream of documents into a corpus you can query.
Automated testing
Agents that exercise a product the way a person would, driving real user journeys and reporting what broke, so coverage does not depend on someone remembering to click through it before a release.
Media understanding
Transcription, translation, summarisation and classification over audio and video at scale, the same engine that powers video analysis inside Vigil.
Evaluation & regression testing
A held-out question set with expected behaviour, run on every change, so you find out that a prompt tweak broke something before your users do.
Private and on-premise deployment
Where a corpus cannot leave the building, we run open-weight models on hardware we build and supply, so nothing is sent to a third-party API.
How we work
How do you keep an LLM system accurate?
Accuracy is an engineering property, not a model property. It comes from controlling what the model can see, constraining what it may do with it, and measuring every change against a fixed evaluation set.
Ground it, then constrain it
Retrieval decides what the model is allowed to see. Prompt engineering decides what it is allowed to do with it. A system that cannot find support for an answer should say so.
Match the register
A correct answer in the wrong voice still fails. We tune output to the writing conventions the audience is graded or judged against, not generic assistant prose.
Measure before and after
Every change is checked against a fixed evaluation set. "It seems better" is not a result, and it is not something you can hand to a stakeholder.
Design for the boring failure
Rate limits, timeouts, malformed tool output, an empty retrieval. Production AI mostly breaks in dull ways, and those are the ones worth engineering against.
Clients
Delivered for
Hwa Chong InstitutionHCI
AI-powered RAG system grounded in the school’s own curriculum and knowledge base, providing H2 Economics students with real-time academic assistance, together with economics-specialised tools like interactive economic graphs to aid explanations.
AngelHack
Developed AI training curriculum for BrainHack TIL-AI 2025, covering the productionisation and deployment with Docker unit of the general AI/ML track and the computer vision track, and provided technical support to 800 competition participants throughout the competition period.
FAQ
LLM & agentic AI: FAQ
- What is retrieval-augmented generation?
- Retrieval-augmented generation, or RAG, is a technique where a language model answers using documents fetched at question time rather than only what it learned during training. The system searches your corpus for relevant passages, puts them in front of the model, and constrains it to answer from that material, which makes answers traceable and far less prone to invention.
- Can retrieval work over images, audio and video as well as text?
- Yes, and that is what multimodal retrieval means: indexing what is inside a file rather than only its title. Speech is transcribed, translated and timestamped, video is described frame by frame, and scanned pages, slides and charts are read with vision models, so all of it is embedded and searchable next to your written documents. A question then returns the paragraph, the slide or the moment in a recording that answers it, with a link back to the original.
- How do you stop an AI assistant from making things up?
- Three things together: retrieval that limits what the model can see to your own vetted material, prompt engineering that instructs it to stay inside that context and decline when support is absent, and an evaluation set run on every change to catch regressions. No approach eliminates error entirely, so we also design the interface so users can check the source.
- Can our data stay inside Singapore, or inside our own network?
- Yes to both. Hosted deployments can be run on servers located in Singapore. Where a corpus cannot leave your premises at all, we deploy open-weight models on hardware we build and supply, so no document is ever sent to a third-party model provider.
- How long does an LLM project take?
- A grounded assistant over an existing document set typically reaches a usable pilot in 1 to 2 weeks, dependent on project scope and the state of the documents. Agentic systems that write to production systems take longer, because the work is in the guardrails and the tool integrations rather than the model.
- Can you integrate with the tools we already use?
- Yes. We have delivered assistants into Google Chat, Slack, WhatsApp, Telegram, web apps, commercial websites and plain HTTP APIs. Meeting users where they already work matters more than another dashboard to remember.
- What does an LLM or agentic AI project cost?
- Quoted per project after scoping. The drivers are corpus size and messiness, how many systems the agent must act against, and whether inference runs on a hosted API or on hardware you own. Running costs are estimated separately and honestly up front, and we accept multiple payment currencies and methods.
Get in touch
Have an unorganised knowledge base no one can use?
That is usually where this starts. Tell us what people keep failing to find.