AI agent development
Agents that do one job well, and ask when they are not sure.
Acasa Labs is an AI agent development company. We build agents that read documents, call your tools, and move work between systems, with every step logged and a person approving what matters. SigmaCPM's tender intake cites a page for every value it extracts, and NoteTube's MCP server lets ChatGPT, Claude, and Cursor work with a student's notes.
The problem
Most agent demos fail the first time they meet real work.
An agent that books, files, or updates things on its own is only useful if it is right, and only safe if you can see what it did. Demos skip both. In production, an agent meets scanned PDFs, missing fields, flaky APIs, and questions it should not answer. The work is in narrowing the job, giving the agent the right tools and data, checking its output, and deciding which steps a person signs off.
What we build
- 01
Agents with a narrow job
One clearly defined task, such as reading a tender, sorting an inbox, or updating a CRM, done reliably rather than everything done loosely.
- 02
Tools the agent can call
Typed functions over your APIs and databases, with permissions per tool, so the agent can act only where you allow it.
- 03
Checks before anything lands
Cited sources, second-pass verification, and rules that drop any output the agent cannot back up.
- 04
People in the loop
Approval queues for the steps that matter, with the evidence beside each item, plus logs and cost for every run.
From production
From AI pipelines we run.
No quote, no value
SigmaCPMEvery value SigmaCPM's model extracts from a tender must come with a quote, and that quote must be found on the cited page. If it is not there, the value is dropped.
A second model checks the first
SigmaCPMBond terms are re-checked by a second, adversarial model run, and disagreements land in a disputed state for a person to resolve. Token use and cost are recorded on every LLM call.
Candidates, not verdicts
Acasa SentinelAcasa Sentinel, our hackathon build, sends watchlist hits to a queue where an operator confirms or rejects each one, with evidence crops and an audit trail.
Your product as an agent's tool
NoteTubeNoteTube's remote MCP server lets ChatGPT, Claude, and Cursor query a student's notes, with OAuth and audience-checked tokens, so an assistant reads only what its user connected.
Honest trade-offs
When you need a workflow, not an agent
If the steps are known and the same every time, write them as a plain workflow with a model call where the input is messy. It is cheaper, faster, and easier to test than an agent deciding what to do next. Use an agent when the path really does depend on what it finds, and even then, keep its tools few and its permissions tight.
What we reach for
- OpenAI, Anthropic, Gemini, OpenRouter
- Python, FastAPI, Django
- Tool calling, MCP
- Postgres job queues, Redis
- OCR, PDF parsing
- Tracing and cost logs
- Evaluation sets from your data
Proof
Shipped, not pitched.
01Our product
SigmaCPM
Construction project management where AI reads the tender and shows its work.
One project record for civil contractors, from tender to closeout: bids, contracts, RFIs, vendor quotes, and drawings. The AI pulls line items and bond terms out of tender PDFs and cites the page for every value.
Web · API · Chrome extension
02Our product
NoteTube
A lecture becomes a study system: chat, flashcards, quizzes, all grounded in the source.
An AI learning platform that turns PDFs, YouTube lectures, recordings, and notes into a structured notebook with grounded chat, spaced-repetition flashcards, and graded quizzes.
Web · iOS · Android · Chrome extension · MCP server
03Case study
Acasa Sentinel
Give it a number plate. It returns the vehicle's route across a CCTV network.
A vendor-neutral video analytics system built for the Gujarat Police Hackathon Innovation Challenge 2026. It reads number plates, re-identifies vehicles by appearance across cameras, reconstructs routes, and raises watchlist alerts for an operator to confirm.
Operator console (web) · API
Questions
AI agent development questions we get asked.
What can an AI agent actually do for our business?
The jobs that need reading and judgement but follow a pattern: pulling data out of documents, sorting requests, drafting replies, updating records across systems, and flagging what needs a person. We start with one of those, not all of them.
How do you stop an agent from making things up?
By making it show its source and checking that source. SigmaCPM drops any extracted value whose quote is not on the cited page, and a second model run re-checks the riskiest fields.
Will the agent act without anyone approving it?
Only where you decide it can. Low-risk steps can run on their own. Anything that sends, pays, or changes a customer record can go to an approval queue first.
Which model do you use?
Whichever does the job best on quality, cost, and data rules. SigmaCPM calls models through OpenRouter, and NoteTube lets each conversation use OpenAI, Gemini, Claude, Llama, or Mistral, so the model can change without rewriting the product.
How do we know the agent is working?
We build an evaluation set from your real cases before launch and measure against it, then log every run in production with its inputs, outputs, and cost.
AI agent development
Which job would you hand to an agent first?
Bring one repetitive task and a few real examples. We will tell you whether an agent or a simpler workflow fits, and what a pilot would look like.