Platform

One platform, end to end. No spreadsheet handoffs.

PrimeTDAP runs the full arc from an engagement session — or a supporting document — to a reviewable deliverable. Six stages, one source of truth, audit-trailed throughout. The deliverable shape adapts to the engagement.

The workflow

Six stages.

Each stage is implemented as a discrete surface in the platform, with its own controls and audit trail.

Ingest

Recorded sessions and supporting documents enter the corpus as cleaned, structured artifacts.

Upload video + transcript (or transcript-only). The platform cleans the VTT, composes a timestamp-anchored markdown transcript, and writes per-encounter metadata.
A heavier reasoning model runs against the cleaned transcript and produces a structured analysis — decisions, action items, risks, themes, people mentioned, sentiment, key quotes.
Supporting documents (SOWs, SOPs, policies, prior reports, contracts) can be uploaded through a separate document surface; ingest, viewing, and download are live, with AI-assisted document analysis available on request.
Re-analyze any session at any time. The new analysis replaces the prior one; the prior version is preserved in the audit log.

Synthesize

Per-session analyses are rolled up into findings, recommendations, and a capability assessment.

Findings are generated per-dimension via the Anthropic Batch API — ~50% cheaper than real-time, with structured strength grades and evidence chains that cite the exact session, speaker, and timestamp.
Recommendations are derived from the findings corpus. Owners are roles, not named individuals. Each carries a horizon (short / medium / long), an execution path (internal / partner / external), and a first action.
A capability assessment classifies the client's in-house strength per dimension as strong, partial, gap, or unknown — surfaced on the Insights deck.

Review

Drafts are reviewed by an engagement professional before promotion. Nothing auto-publishes.

Every finding and recommendation moves through a status state machine: draft → reviewed → Counsel-approved. Promotion requires a deliberate action.
Re-syncs preserve any row already at reviewed or approved status. The platform never silently overwrites human-reviewed work.
A per-call AI audit log records the model, prompt version, token counts, cost, and outcome of every inference — visible to administrators.

Package

Slide-summary decks and a formal memorandum, rendered server-side.

Five decks are produced: Executive Summary, Findings, Insights, Recommendations, and a restricted CEO Readout. All rendered as HTML with browser-print to PDF — no PowerPoint dependency.
A formal memorandum is generated as both Word (DOCX) and PDF — cover, privilege framing, table of contents, executive summary, methodology, per-dimension intros, per-horizon recommendation intros, conclusion, appendices.
Generated content is cached in the database; regeneration is a one-click admin action that completes in under a minute for the memorandum.

Distribute

Client leadership accesses the work product through a tier-gated, watermarked view.

Client users sign in with a one-time code sent to email (SMS optional). Their view is read-only, watermarked, and framed with "do not forward" language on every page.
A separate admin tier — Primus engagement staff — sees the full workspace including transcripts, the AI audit log, security events, and platform administration.
The CEO Readout deck has a fourth, deck-specific allowlist — only the engagement leads, not the broader audience.

Interact

A bounded chat per scope — per-person, per-session, or per-meeting — with citation discipline.

Chat is scoped tightly. Asking about a person retrieves only that person's sessions; asking about a meeting retrieves only that meeting. There is no cross-corpus chat.
Every answer cites the source — timestamp, speaker, evidence chain — so a claim can be verified against the underlying recording or analysis.
Off-topic questions are politely declined. Rate limits and per-user daily spend caps apply. Every call is logged and cost-attributed.

AI, disciplined

One provider. Three model tiers. Disciplined surfaces.

All AI inference goes through Anthropic Claude. Every system prompt is hashed and version-tracked; every call is cost-attributed; every output cites its evidence. Specific model versions are listed in the AI Disclosure / Model Card.

Analytical synthesis

Heavier reasoning tier

Per-session analysis (decisions, risks, action items, themes)
Per-dimension findings (cost-efficient batch processing)
Recommendations generation
Capability assessment
Per-person briefings
Meeting minutes & notes

Structured rendering

High-volume tier

Slide summaries (per finding, per recommendation)
Executive Summary deck content
Formal memorandum prose blocks

Bounded chat

High-volume tier

Per-person chat (scoped to one person's sessions)
Per-meeting chat (scoped to one meeting)
Citation discipline + scope refusal + injection defence

Citation discipline

Chat answers cite the timestamp, speaker, and analysis they draw from. Findings cite the encounter and evidence chain. The memorandum cites findings by identifier.

Scope discipline

Each chat is bounded — to a person, a session, or a meeting. Off-topic questions are politely declined and the refusal is logged as a security event.

Security discipline

System prompts treat retrieved content as data, not instructions. Meta-instructions ("ignore previous", "reveal system prompt") are refused. The system prompt is never disclosed.

Prompt versioning

Every prompt is SHA-256 hashed and registered at startup. The forensic answer to "what prompt did the model see on day X?" is one SQL query.

Fenced content: sensitive transcripts can be marked as fenced — they never enter the AI context, never appear in any chat, and never leave the originating workstation.

Deliverables

The artifact at the end shifts to match the engagement.

Counsel-directed assessments produce a formal memorandum and a deck set. Audit-readiness engagements produce an artifact inventory, a strengthened submission package, and interview-prep briefs. Due-diligence engagements produce a workstream-organised diligence report and a materiality-graded issue list. Same arc, different shape.

For Counsel-directed assessment engagements

Deliverable	Detail
Formal memorandum (DOCX + PDF)	Cover, privilege framing, table of contents, executive summary, methodology, per-dimension intros, per-horizon recommendation intros, conclusion, appendices.
Executive Summary deck	Thesis, scorecard, insights triangulation, stakes, and program-marker arc.
Findings deck	Per-dimension drill-down with evidence chains, strength grades, and source citations.
Insights deck	Capability heat-map across dimensions plus the recommendation shape across horizons.
Recommendations deck	Per-horizon detail: title, key statistic, bullets, owner role, execution path, first action, success signal.
CEO Readout deck (restricted)	Allowlist-gated. Engagement leads only.

All decks render as HTML with browser print-to-PDF. The memorandum renders as both Word (python-docx) and PDF (reportlab) so it slots cleanly into Counsel's existing file conventions.

For audit & external-review readiness engagements

Deliverable	Detail
Artifact-request inventory	One row per request from the external reviewer: what they asked for, what exists today, where the gap is, who owns the fix, target submission date.
Strengthened submission package	Documents organised, gaps closed, formatting consistent, lifecycle completeness verified. Ready to send.
Interview-prep briefs	Per-respondent role-based briefing: what the reviewer is likely to ask, what supporting context to have ready, what not to volunteer.
Cycle-support advisory notes	Real-time notes through the review cycle: interpretation of what is being examined, consistency checks between submitted artifacts and interview responses, supplemental-request management.

For due-diligence engagements

The same arc extends to diligence engagements; the deliverable shape below is what the platform can produce when applied to that use case. Configured at engagement onboarding.

Deliverable	Detail
Diligence report	Executive summary, per-workstream findings (commercial, operational, technology, financial, legal, HR), risk register, conclusion. Renders as Word and PDF for deal-team distribution.
Materiality-graded issue register	Open issues from the diligence corpus with materiality tier, owner, current status, target resolution. Filters by workstream or by deal-blocking severity.
Management Q&A trail	Every question raised to management, who asked, when answered, with the supporting evidence cited. Defensible record of the diligence conversation.
Red-flag executive brief	Top issues by materiality, deal-team facing, drives go / no-go and price-adjustment conversations.

All three deliverable shapes draw from the same underlying corpus and the same six-stage workflow. The platform doesn't re-do the work for each engagement type — it re-points it.

Access

Three tiers. Strict separation.

Tier is checked at the route, at the template, and at the build — defence in depth across the rendering pipeline.

Primus internal

Engagement administrator

Full read/write. Ingest sessions, generate AI artifacts, edit findings and recommendations, promote drafts, generate the memorandum, manage users, view the audit log.

Primus internal

Engagement viewer

Read everything Primus-internal — findings, recommendations, internal levers — with no edit, generate, or delete permissions.

Client leadership

Client viewer

Read-only access to findings, recommendations, decks, and the memorandum. Watermarked. "Do not forward" framing. No transcripts, no internal levers, no audit-log surfaces.

Authentication: Primus team via Microsoft Entra single sign-on with tenant MFA; client users via email one-time codes (SMS optional). Per-tenant SSO configurations available on request.

Honest scope

What PrimeTDAP is not.

Calling out what the platform doesn't do is as important as describing what it does — so the fit is clear from the first conversation.

Not a decision-making AI. Every AI artifact is reviewed before promotion. The platform produces decision-support, not decisions.
Not a recording editor. Recordings are read-only references; the platform doesn't transcode, splice, or modify them.
Not a legal-advice tool. The platform produces technical and operational work product. Legal positioning is added by the engaging customer's legal counsel where applicable.
Not a self-serve SaaS. Single-tenant per engagement. Onboarding is Primus-led — we provision your environment, configure access, and walk your team through the workflow before the first session.
Not a public-facing analytics surface. Every analytical page is auth-gated; only the policy documents are public.

See it on your own engagement content.

A demo runs against a sanitized fixture set. Onboarding to a real engagement happens after MSA + DPA are countersigned.

Request a demo Read the security posture