Talk to us
Menu

Past work

Two products that we built and retired.

Before the studio, we built our own AI products. We took two of them to a private beta, and we retired both in 2026. We do not sell them. They show how we build: the same approval steps, tests, and limits go into each client agent.

Applicant tracking with an AI interviewer

VikaHire

Retired

VikaHire was an applicant tracking system for recruiters, hiring managers, and interviewers. AI did the routine work, and a person on the hiring team made each decision.

What we built

An AI interviewer
It ran voice rounds, coding rounds, and system-design rounds, and asked follow-up questions. It never ran the code of a candidate.
A person made the decision
AI scores were advisory. After an AI interview, a candidate moved forward only when a person recorded the decision.
Approval before change
The assistant proposed a change. A person confirmed the change before it ran.
Input stayed data
Résumés and answers reached the model as untrusted data. Each AI interview score cited evidence in the transcript.
Usage and limits for each organization
VikaHire recorded each AI call without its content. A limit warned at 80% and stopped new AI interviews at 100%.
Isolation and audit
Each database query stayed inside one organization. An append-only log recorded approvals and agent changes.
The VikaHire pipeline board. Columns show each hiring stage, from New to Hired, with candidate cards in the first two stages.
A VikaHire pipeline with seeded demo data.
A VikaHire AI interview rubric. Each item shows its score, weight, guidance, written reason, evidence control, and override control.
An AI interview review with seeded demo data.
VikaHire agent settings for concurrency, open runs, step timeout, token budget, and run deadline.
Agent runtime limits with seeded demo data.

Range-based project estimation

VikaEstimate

Retired

VikaEstimate turned a brief into tasks, hours, roles, and costs. Each task carried a confidence range, so a team saw the margin before it sent a price.

What we built

Ranges, not padding
Confidence factors showed where unknowns changed the hours. PERT gave the entered hours four times the weight of each bound.
Margin while you edit
Cost, revenue, and the low, mid, and high margin moved with each task.
AI grounded in delivered work
The AI proposed sections, tasks, hours, roles, and assumptions. Won projects gave it context, and recorded delivery hours showed estimation bias by role.
One calculation everywhere
The browser and the API called the same shared functions. Each project kept a frozen copy of its rate card.
The VikaEstimate workspace. Sections contain tasks with hours, roles, confidence factors, range bars, and revenue ranges.
The estimate workspace with seeded demo data.
A VikaEstimate project summary with range bars for effort, cost, and revenue, plus the mid margin and project details.
A project summary with seeded demo data.

Want this kind of engineering in your agent?

Tell us about the task. We will tell you whether an agent fits.

Talk to us