Research
Research we can ship.
We work on applied AI, language tools and the infrastructure under our products. What proves useful goes into those products and out as open-source packages anyone can install.
- Focus
- Applied AI, language tools, developer infrastructure
- Output
- Open-source packages, used in our own products first
- Catalogue
- More than 200 packages
- Not doing
- Robotics and space are direction, not active work
What we work on
Six areas, each with code you can read and install today.
AI building blocks
One way to talk to any model: chat, streaming, prompts, token budgets, tool definitions and repairing the JSON models return.
Retrieval & agents
Splitting, embedding, storing and reranking documents, and a small agent loop that can call your own tools.
Cheaper AI
Spending fewer tokens: cheap checks before a model call, caching repeated calls, condensing sources and failing over between providers.
Language & dates
Nepali and English names across scripts, Bikram Sambat dates, and reading when an article was really published.
Trust & checks
Redacting personal data, catching prompt injection, and checking that numbers and names in generated text match the source.
Reliable software
Zero-dependency libraries and a testing kit, so what we publish stays small, auditable and easy to keep up to date.
The work, as packages
A selection from the catalogue, grouped by area. Each one is zero-dependency and documented on developer.lacspace.com.
31 packages
@lacspace/ai1.2.1One chat and stream API over fetch for OpenAI, Anthropic, Gemini and compatible models.AI building blocks@lacspace/prompt1.1.0Build, compose and render prompts with typed variables.AI building blocks@lacspace/tokenizer1.1.0Token estimates, cost and context-budget management.AI building blocks@lacspace/json-repair1.1.0Extract and repair JSON from messy model output.AI building blocks@lacspace/stream1.1.0Parse Server-Sent Events, NDJSON and streaming model replies.AI building blocks@lacspace/ai-tools1.1.0Define a tool once, get the spec for any provider.AI building blocks@lacspace/chunk1.1.0Split text, Markdown and code for retrieval and long prompts.Retrieval & agents@lacspace/embeddings1.0.0Provider-agnostic embeddings client plus vector maths.Retrieval & agents@lacspace/vector1.0.0In-memory vector store for semantic search.Retrieval & agents@lacspace/rag1.0.0Index documents, retrieve and assemble prompt context.Retrieval & agents@lacspace/rerank1.0.0BM25, TF-IDF, rank fusion and hybrid reranking.Retrieval & agents@lacspace/agent1.0.0A small tool-calling agent loop for any model and any tools.Retrieval & agents@lacspace/memory1.0.0Conversation memory kept inside a token budget.Retrieval & agents@lacspace/providers1.0.0Connection presets for model and embedding endpoints.Retrieval & agents@lacspace/condense1.0.1Turn several sources on one story into one short digest that keeps the numbers.Cheaper AI@lacspace/screen1.1.0A cheap lexical gate that decides what is worth a model call.Cheaper AI@lacspace/keyphrase1.0.0Keyphrases, tags and entities without a model, in English and Nepali.Cheaper AI@lacspace/llm-cache1.0.0Cache identical model calls by model, prompt version and input.Cheaper AI@lacspace/keypool1.4.0Fail over across providers and pool keys so free tiers go further.Cheaper AI@lacspace/translit1.0.5Nepali and English name transliteration and fuzzy matching.Language & dates@lacspace/nepali-date1.2.1Bikram Sambat and Gregorian conversion with Nepali formatting.Language & dates@lacspace/datecheck1.0.1Find when an article was really published, including Nepali dates.Language & dates@lacspace/extractive1.2.2An extractive brief of an article: summary, key facts and entities.Language & dates@lacspace/trend-detect1.0.0Spot terms that are suddenly trending in your own stream of items.Language & dates@lacspace/moderation1.0.0Find and redact personal data, flag unsafe text and prompt injection.Trust & checks@lacspace/eval1.0.0Score model output with deterministic checks.Trust & checks@lacspace/factcheck-lite1.0.1Check that every number, date and name in generated text matches the source.Trust & checks@lacspace/expect1.1.0A fluent assertion library that works in any test runner.Testing@lacspace/spy1.1.0Spies, stubs, mocks and fake timers.Testing@lacspace/fixtures1.1.0Deterministic, typed test data from a seed.Testing@lacspace/snapshot1.1.0Stable value serialising and snapshot testing.Testing
Where it runs
The same pieces, working inside things people use.
- Lacspace Mail
- Lumi AI summaries, drafts in your tone, translation and a pre-send review; phishing checks on SPF, DKIM and DMARC; smart categories.
- lacspace.com
- Lumi, the assistant on this site, answers questions about Lacspace and its products.
- StockYatra
- Lumi AI inside the trading simulator, with more of it on the Pro plan.
- Newsroom tools
- Briefs, fact-consistency checks, date checks and name matching built for an AI-assisted newsroom, then published as packages.
How we approach it
- Cheap checks first
- A lexical screen, a cache or a deterministic rule runs before any model call. The model handles only what is left.
- Bring your own model
- Our AI packages hold no keys and favour no provider. You choose the model, and can change it later.
- Zero dependencies
- Each package stands alone, so installs stay small and there is less to audit or patch.
- Publish what we use
- If we build a piece for our own products and it is useful elsewhere, it goes out as an open-source package.
- Measure the output
- Generated text is scored and checked against its source rather than trusted because it reads well.
Before we build a product
Product research is less glamorous: three questions answered before a timeline is promised.
- 01
Who it is for
Interviews and walkthroughs with the people who will use it day to day.
- 02
What already exists
A review of competing products, their pricing and where their users complain.
- 03
Is it buildable
A technical spike on the risky parts before a timeline is promised.
Writing
- Why zero-dependency npm packages: faster installs, smaller bundles, fewer CVEs
- Your seven security layers, as npm packages
- SEO for humans and AI: sitemap.xml, robots.txt, llms.txt and JSON-LD as code
- The Lacspace developer handbook: every package, cross-linked
- How to prepare your business for the AI era
- Black-box AI trading: how it works and the red flags to avoid
Where it is heading
Software, then AI, then robotics, then space. The first two are today's work; the last two are direction, not programmes.
- Today
Software
Products and client systems, and the open-source tools under them.
- Today
AI
Applied AI inside those products, and the packages on this page.
- Direction
Robotics
Where we want the work to lead. Not an active programme today.
- Direction
Space
The long horizon. Not an active programme today.
Building with AI, or want to work on this?
Use the packages, tell us what is missing, or come and build them with us.