Paste a job description and your resume. Lattice understands the role, reads your evidence, scores your hiring readiness from three independent reviewers, and interviews you out loud — then remembers what to fix next time.
Built on a production-grade,
multi-agent stack.




The Career Orchestrator turns your goal into a plan and runs independent work in parallel — concurrent agents, not a sequential checklist.
Every skill in your profile carries the sentence and section it came from. Anything unsupported gets flagged as missing evidence rather than quietly asserted.
Three independent specialists score you concurrently on different dimensions. When they disagree, the outlier is re-run once with the others’ reasoning before anything is finalized.
Live voice interviews are transcribed in full. Evaluation is derived from the transcript by a dedicated specialist — never a hardcoded rating.
Confirmed weaknesses from each interview roll forward into your candidate profile — reviewable, and never written without an evaluation behind it.
Your resume is chunked, embedded locally and retrieved per question, so the chat answers only from your own evidence.
Your resume shows backend experience through Django API development and PostgreSQL data pipelines.
ATS, Recruiter, and Hiring Manager agents score you concurrently, not in sequence. Nothing is scored without a stated reason — disagreements are shown, not averaged away. A dedicated Job Match node scores 11 weighted dimensions against this exact posting, alongside it.
Gemini Live or Pipecat on the other side — interrupt mid-sentence and it responds naturally — a plan weighted to the role, and evaluation derived from what you actually said, not a template.
Confirmed weaknesses roll forward. Reviewable, and never written without an evaluation behind it.
Every read is live. Your real signals, real restaurants, a real cart. Every write stops and asks you first.
Reads yesterday from your own data — readiness moves, deadlines closing, roles worth a look — and writes the one paragraph that matters.
Ask in plain language. It searches live restaurants and real menus through MCP, then proposes an order with actual prices.
Or skip the chat: real menus, a veg filter, a live Swiggy cart with the server’s own fees and taxes — kept in its own tab.
The cart and the payment methods are genuinely yours. Placing the order is always simulated — no restaurant is ever sent anything.
Because the second a machine can apply for you, you inherit everything it got wrong, at the speed it can send. A bad match, a hallucinated claim, a recruiter who now remembers your name for the wrong reason.
Every write pauses the graph on a real interrupt — not a confirm dialog bolted on afterwards. Approve it, or reject it with a reason the agent reads before it tries again. There are eight of these moments, and none of them is “submit application”.
Draft ready — 3 leads, public context only. No warm intro is implied.
Outreach is drafted, never dispatched: there is no send path in the codebase to begin with. You copy the message, send it from your own account, and mark it sent — so the record matches what actually happened.