Five working demos from the lab — exception triage, reconciliation, spec writing, meeting intelligence, and job-fit analysis. Every demo runs entirely in your browser; nothing you type ever leaves your device.
How to read these demos: each agent is a real working prototype — a rule-based reasoning engine running 100% client-side. What you type in is actually processed; outputs change with your inputs. They show our product thinking: problem framing, UX, and output design.
Below each demo, the "Production path" notes describe how we'd rebuild it with LLMs, retrieval, and evals. All sample data is fictional.
Paste a description of an operations exception — a break, a fail, a mismatch — and the agent classifies it, scores its confidence, assesses severity, hypothesizes root causes, and drafts a remediation ticket.
Run the demo — or load a sample — to see the agent work.
Runs 100% in your browser — nothing you type leaves your device.
A weighted keyword classifier: multi-word phrases score 3 points, single terms score 1. Scores are normalized into a confidence value, severity comes from urgency and aging signals in the text, and root causes, next steps, and the ticket are assembled from per-category operations playbooks.
Swap the keyword weights for a fine-tuned classifier trained on labeled break tickets, add retrieval over the firm's runbook library for root-cause hypotheses, and measure precision/recall per category before wiring it into the ticketing system.
Paste two position snapshots — expected vs. actual — and the assistant matches them line by line, flags every break, and explains each one in plain English with a likely cause.
Run the demo — or load a sample — to see the agent work.
Runs 100% in your browser — nothing you type leaves your device.
A tolerant CSV parser normalizes each row, positions are matched by ID, then deterministic break rules fire: missing, unexpected, quantity mismatch, and price variance over 0.5%. Each break gets a plain-English explanation from a template keyed to its type.
Connect directly to the accounting and custody feeds, learn variance thresholds per security from history, and add an LLM layer that drafts analyst-ready break commentary from the structured findings. Evals: break recall measured against the ops team's manual review.
Describe an operations workflow in plain English and the agent drafts a mini-spec: numbered business requirements, user stories per actor, and Given/When/Then acceptance criteria.
Run the demo — or load a sample — to see the agent work.
Runs 100% in your browser — nothing you type leaves your device.
The description is split into sentences — one business requirement each — then into clauses filtered by an action-verb heuristic. Each actor gets user stories built from the strongest clauses, and the first stories get Given/When/Then acceptance criteria generated from templates.
Use an LLM with retrieval over the existing spec library for consistent terminology, generate edge cases and non-functional requirements, and score drafts against a rubric — testability, completeness, ambiguity — reviewed by a business analyst before sign-off.
Paste raw meeting notes and the agent pulls out action items (with owners and due dates), risks and blockers, and decisions — the follow-up email, drafted for you.
Run the demo — or load a sample — to see the agent work.
Runs 100% in your browser — nothing you type leaves your device.
Line-by-line pattern matching: action prefixes (TODO, Action, AI), commitment-plus-deadline phrasing, and "Name to verb" assignment patterns become actions with extracted owners and dates. Separate patterns catch risks, blockers, and decisions.
Pipe meeting transcripts through diarization plus an LLM extractor constrained to a strict JSON schema, resolve owners against the company directory, and sync actions straight to the task tracker. Evals: action-extraction F1 against human-annotated meetings.
Paste any job description and the analyst scores it against a finance-operations product profile — matched skills, gaps to close, and ready-made talking points for the interview.
Run the demo — or load a sample — to see the agent work.
Runs 100% in your browser — nothing you type leaves your device.
The description is scanned against a 60-skill lexicon plus a finance-domain subset; seniority comes from the years-required pattern. The score blends skill overlap with the target profile, domain density, and seniority fit — gaps and talking points are derived from the same matches.
Replace substring matching with embedding-based skill extraction, calibrate the scoring model against real hiring outcomes, and add retrieval over verified experience for evidence-backed talking points instead of templated ones.
These are prototypes. We build the production versions — scoped to your workflows, integrated with your systems, and measured against your KPIs.