Tengwar
Practice Areas · Vol. I

Two theaters, one firm.

We deploy inside portfolio companies and inside the sponsor itself. The same engineering discipline — forward-deployed, model-neutral, eval-driven — applied to both sides of the table.

Theater One

Portfolio companies.

Six practices inside the operating company — where frontier models earn their keep against the KPIs the sponsor already tracks.

I.
Operations & Supply Chain

The oldest levers in a PE playbook, now redrawn by models that can read a bill of lading and a contract in the same pass. We rebuild forecasting, inventory positioning, procurement, and exception handling around systems that surface decisions rather than dashboards. The work is concrete and the KPIs are the ones the CFO already tracks.

Typical engagement shape
Embedded team of three to five. Twelve to twenty-four weeks. Focus on one or two high-volume decision loops, not a horizontal transformation.
Common outcomes
Cycle time compression on purchase orders and exceptions. Working capital release from better demand signals. A durable internal team that owns the stack after handoff.
Tools and models
Claude, GPT, Gemini, open-weight Llama variants for on-prem inference, plus orthogonal tooling for document extraction and evaluation.
II.
Revenue & GTM

Pipelines are the obvious target and the one most engagements get wrong. The value is not in writing better outbound copy; it is in re-architecting how a portco qualifies, routes, and prices an inbound request against the specific contours of its vertical. We build the supporting infrastructure first and the agent-style interfaces second.

Typical engagement shape
Embedded team of two to four. Eight to sixteen weeks. Co-owned with the portco revenue leader and a single engineering counterpart.
Common outcomes
Measurable lift on conversion and time-to-quote. Pricing that reflects the actual shape of demand. Reps spending more time in front of customers.
Tools and models
Claude, GPT, vertical embeddings, and evaluation harnesses tied to conversion metrics rather than sentiment scores.
III.
Finance & Back Office

Close cycles, reconciliation, vendor onboarding, audit preparation — the mechanical work that sets the pace of every portco month-end. Frontier models handle these loops reliably when the surrounding pipeline is built properly. We build the pipeline first and let the models do the work they are actually good at.

Typical engagement shape
Embedded team of two to three, often sitting alongside the controller. Eight to twelve weeks for a first production loop, with a second loop staged immediately after.
Common outcomes
Days shaved off the close. Reconciliation work moving from manual to exception-review. Audit trails that a Big Four auditor will sign.
Tools and models
Claude for long-context document reasoning, GPT for structured extraction, open-weight models where data cannot leave the environment.
IV.
Customer Experience

Contact centers, self-serve portals, and the long tail of account management. The work is not about replacing humans; it is about moving the boundary between a supervised response and an unsupervised one deliberately, with evaluation pipelines that make the boundary visible to the operating committee.

Typical engagement shape
Embedded team of three to five, with a dedicated evaluation engineer. Twelve to twenty-four weeks. Tight feedback loop with the portco CX leadership.
Common outcomes
Handle-time compression. Containment lift on well-understood intents. A tested failure-mode catalogue the portco can defend.
Tools and models
Claude, Gemini for multimodal transcript analysis, GPT for tool-use orchestration, and an evaluation harness bespoke to the portco's intents.
V.
Knowledge Work Automation

Underwriting memos, research notes, investment committee packs, engineering reviews, policy and claims packets — the knowledge work that a portco's senior people spend half their week producing. We target the loops where the format is stable, the inputs are machine-readable, and the reviewer is still a human with a pen.

Typical engagement shape
Embedded team of two to four. Ten to sixteen weeks. Scoped around a single document type and a single approving reviewer.
Common outcomes
Time saved per memo. More consistent output. Senior reviewers spending their time on judgment instead of formatting.
Tools and models
Claude Opus for long-context drafting, GPT for structured review, Llama for private-data processing, plus purpose-built evaluation.
VI.
Risk & Compliance

Model-assisted review of contracts, policies, filings, and transaction data. This is the practice area with the sharpest regulatory edges and the one where model neutrality matters most — the regulated entity cannot afford to be locked into a single vendor when the supervisor asks a pointed question in eighteen months.

Typical engagement shape
Embedded team of three to five, including a compliance-fluent engineer. Twelve to twenty-four weeks. Governed by the portco's existing risk committee.
Common outcomes
Coverage lift against a defined risk taxonomy. Auditable model decisions. Documentation a regulator will accept.
Tools and models
Claude, GPT, and open-weight models where required by policy, with evaluation pipelines designed to survive a regulatory exam.
Theater Two

The sponsor itself.

Every function inside a PE firm is knowledge work performed by expensive humans reading documents and producing documents. The entire operation is LLM-shaped. We deploy inside the firm — deal team, finance, IR, legal — with the same engineering discipline.

I.
Deal Origination & Sourcing

Associates screen hundreds of CIMs and broker teasers each year. Frontier models read all of them in hours, score each against the fund's investment criteria, surface the twenty that matter, and draft preliminary memos with citations to the source material. What took a team a week takes a pipeline an afternoon.

Typical engagement shape
Two to three engineers embedded with the deal team. Eight to twelve weeks to production. Ongoing model tuning as the fund's thesis evolves.
Common outcomes
Higher deal-flow throughput with fewer passes. Fewer missed opportunities in the long tail. Sourcing hypotheses generated proactively rather than reactively.
Tools and models
Claude Opus for long-context CIM analysis, GPT for structured extraction, custom scoring models fine-tuned on the fund's historical deal data.
II.
Due Diligence

The data room is the most document-intensive phase of any deal — thousands of contracts, financial statements, org charts, customer data, and legal filings, read under time pressure by expensive humans. AI reads the entire room in hours, flags risks, extracts key terms from every contract, cross-references customer concentration, and drafts the diligence tracker automatically. The deal team reviews exceptions, not documents.

Typical engagement shape
Two to four engineers deployed alongside the deal team for the duration of the diligence process. Typically six to ten weeks, with deliverables staging alongside the deal timeline.
Common outcomes
Diligence timelines compressed by weeks, not days. Risk flags surfaced earlier. IC memo first-draft quality from structured outputs rather than manual synthesis.
Tools and models
Claude Opus for contract and document reasoning, Gemini for multimodal analysis of scanned materials, GPT for structured data extraction, Llama for on-prem processing when data room terms restrict cloud inference.
III.
Financial Modeling & IC Execution

LBO models, sensitivity tables, scenario analysis, returns waterfalls — the mechanical core of deal execution. AI builds the first-pass model from the CIM and data room outputs, runs fifty scenarios instead of five, audits formulas, and generates sensitivity tables. IC memos that take two to three weeks of prose synthesis come back as structured first drafts in forty-eight hours. The team adds judgment, not formatting.

Typical engagement shape
One to two engineers embedded with the deal team's modelers and associates. Six to twelve weeks. Scoped around the fund's existing model templates and IC memo format.
Common outcomes
Modeling turnaround from days to hours. IC memo cycles shortened. More scenario coverage per deal. Fewer formula errors.
Tools and models
Claude for long-form memo drafting and model logic, GPT for spreadsheet formula generation and audit, custom evaluation pipelines benchmarked against the fund's own historical models.
IV.
Portfolio Monitoring & Operating Committee

Operating teams manually collect monthly KPIs from thirty to eighty portfolio companies via email and spreadsheet, build dashboards, and write operating committee memos. AI agents pull data directly from portco systems or standardized reporting templates, flag variances against plan, generate the operating committee memo with commentary, and build the dashboard. The operating partner reads a finished analysis, not raw data.

Typical engagement shape
Two to three engineers working with the portfolio operations team. Twelve to sixteen weeks for full pipeline deployment across the portfolio. Ongoing tuning as portcos are added or exited.
Common outcomes
Operating committee prep time cut from days to hours. Variance commentary generated automatically. Earlier detection of underperformance across the portfolio.
Tools and models
Claude for memo generation and variance narrative, GPT for structured data ingestion and anomaly detection, custom integrations with portco ERP and reporting systems.
V.
LP Reporting & Investor Relations

Quarterly letters, DDQs, annual meeting materials, capital call notices, performance attribution — the IR function is almost entirely document production against structured data. The quarterly letter drafts itself from fund performance data. The DDQ engine maps your canonical answers to every new questionnaire format, flagging only the questions that need fresh human input. Annual meeting decks assemble automatically from quarterly outputs.

Typical engagement shape
Two engineers embedded with the IR and fund finance teams. Eight to twelve weeks for the initial pipeline. Quarterly cadence reviews thereafter.
Common outcomes
DDQ response time from weeks to days. Quarterly letter drafts produced in hours. Consistent messaging across LP communications. IR team capacity freed for relationship management rather than document production.
Tools and models
Claude for long-form letter drafting and DDQ mapping, GPT for data extraction and formatting, evaluation harnesses that check consistency across all LP-facing materials.
VI.
Fund Finance, Legal & Compliance

NAV calculations, waterfall computations, management fee tracking, carried interest allocation, audit prep, side letter tracking, regulatory filings. The mechanical backbone of fund operations, performed by senior professionals whose time is better spent on judgment calls. AI runs the waterfall, reconciles NAV, prepares audit workpapers, tracks side letter obligations across LPs, drafts Form PF and ADV filings from structured data, and flags discrepancies before the CFO or GC sees them.

Typical engagement shape
Two to three engineers embedded with fund finance and legal. Twelve to twenty-four weeks given the compliance sensitivity. Governed by the fund's existing audit and compliance committees.
Common outcomes
NAV reconciliation automated. Waterfall computations auditable and instant. Side letter obligations tracked systematically rather than in spreadsheets. Regulatory filing prep time compressed.
Tools and models
Claude for document reasoning and filing drafts, GPT for structured computation and reconciliation, open-weight models where fund terms restrict cloud processing, evaluation pipelines designed for audit-trail requirements.