Bench Starter
The exact agentic-engineering setup we install for clients, as a public repo with a live eval and cost dashboard: fork it, or hire us to install it.
Five run today, on synthetic data, with their eval scores, cost meters and traces on the page. Eighteen more are proposed across eleven domains, scored on business appeal, differentiation, effort, cost and coverage. Nothing is marked live before it runs.
Chosen because each one proves the setup every other demo is built on: the bench, the gateway, the eval gate and the trace.
The exact agentic-engineering setup we install for clients, as a public repo with a live eval and cost dashboard: fork it, or hire us to install it.
Six real ways agents fail in production (loops, cost blow-ups, tool misuse, stale retrieval, prompt injection, silent regressions), replayed from traces, with the fix for each.
Ask anything about the employee handbook and leave policy; get an answer with the exact clause cited, or an honest "not in the policy".
Drop in a month of purchase invoices and your GSTR-2B; see every mismatch explained and the vendor emails drafted.
Upload an event log from your ERP; see the real process, where it leaks time, and the three AI opportunities with a business case.
Every one is planned. Ask for the one that maps to your workflow and it moves up the list.
Send one prompt to three models through our cost gateway and watch price, latency, cache hits and the budget cap behave in real time.
Paste a project plan; get the ten ways it fails, ranked, with the mitigations, using the same skills we use on client work.
Describe your business in five lines; get your three best AI opportunities, the phase you are at, and what the first step costs.
Watch a pull request to an AI feature get blocked because the answer quality dropped, then pass after the fix.
Ask "why did line 3 keep stopping last week?" and get an answer that connects shift notes, maintenance logs and the asset registry.
Ask a plain question of last quarter's numbers; get the chart, the query it ran, and a one-page memo, in one minute.
A WhatsApp-style assistant that answers a farmer's price and crop questions in Hindi, Marathi or Telugu, from public data, for under one rupee a conversation.
Complaints in any language get classified, routed to the right department, acknowledged within the SLA, and every decision is explainable.
Paste a GST or income-tax notice; get a structured reply draft citing the circulars, with the partner approval step built in.
A firm's memos become an assistant that answers like the senior partner, says when it is unsure, and shows what each answer cost.
Open a 40-page commercial contract; get the 41 clauses lawyers look for, flagged and scored against an expert-labelled benchmark.
Turn a year of safety observations and near-miss reports into trends, root-cause themes and next week's toolbox talk.
Paste a month of reviews; see what customers actually complain about, by aspect, and the replies drafted.
An agent that reads shipment exception emails, checks the TMS, proposes the fix, and waits for a human before it acts.
Turn interview notes into a structured, evidence-quoted scorecard, with the bias guard on.
Patient messages in, structured triage summaries out, with hard rules that the model cannot cross.
Turn 50 product photos and thin titles into complete, consistent listings in three tones, and see what the batch cost.
Photograph a leaf; get a likely disease, a confidence, and advice grounded in public advisories, or an honest "not sure, ask an agronomist".