The opportunity
The $7B PE firm needed answers that reached across its largest data stores.
The firm wanted to know two things: What changed across the portfolio this quarter? Who in its network could make a credible introduction to a company under evaluation? Answering either question required more than a quick search.
01
Portfolio monitoring
Build one complete view of portfolio performance.
The agent had to gather and organize eight quarters of financial data for 14 companies without leaving a company or reporting period out.
02
Network intelligence
Find the strongest route to the right person.
The agent had to resolve target companies, work through ambiguous matches, trace relationships, and rank the best introduction paths.
112
company-quarter comparisons in each portfolio update.
150,000+
company records in the firm’s CRM.
<60 sec
to return a relationship brief for a target.
How the architecture reduced hallucinations
Portfolio monitoring
No hallucinated answers across 100+ evaluations, at $0.03 per complete portfolio update.
The earlier portfolio dashboard could not update itself reliably. When records were missed or too much raw data accumulated, the model could stop early or fill gaps with invented figures. Port of Context rebuilt the workflow to gather every required record, standardize 112 company-quarter comparisons, and return one complete eight-quarter view across all 14 companies.
Collect
Include every company and quarter.
The research layer works through the full set of records required for the update.
Standardize
Make the figures comparable.
Financial data is organized into the same fields across companies and reporting periods.
Deliver
Return one complete portfolio view.
The agent receives the finished eight-quarter summary and presents it to the team.
The same design that kept hallucinations out of the evaluations also reduced token waste. Port of Context checks and organizes the 14-company dataset before the model prepares the update, so the model works from verified evidence instead of carrying page after page of raw financial data.
Specialized architecture
$0.03
measured input cost for the complete 14-company update.
Conventional workflow
$46.38
estimated input cost when pages and intermediate results pass through the model.
99.9% lower model input cost while preserving the complete portfolio view.
Network intelligence
The firm’s 150,000+ company CRM could return a relationship brief in under 60 seconds.
The firm’s CRM and relationship history had accumulated over decades. For six target companies, the agent identified the correct records, worked through ambiguous matches, traced paths across people and prior investments, and ranked the strongest routes. The deal team received the best-supported paths with recency and relationship-strength signals.
Resolve
Identify the correct companies.
Ambiguous names and incomplete records are checked before the relationship search begins.
Trace
Follow the credible paths.
The research layer works across people, companies, and prior relationships.
Rank
Return the strongest routes.
The agent receives the best-supported matches and recommends where the team should start.
Making the relationship search economical to repeat
The firm’s target list and relationship data change throughout the deal cycle. To keep each brief current, the team needs to rerun the entire search, including entity checks, fallbacks, relationship tracing, and ranking. Port of Context completed that six-target workflow for $0.17 in model input, compared with an estimated $1.97 when the intermediate search results pass through the model.
Specialized architecture
$0.17
measured model input cost for the complete six-target search.
Conventional workflow
$1.97
estimated model input cost for the same searches when intermediate results pass through the model.
91% lower model input cost for the same complete six-target workflow.
The result
The firm gained reliable production workflows across both use cases.
Across 100+ portfolio-monitoring evaluations, the production workflow returned no hallucinated answers. The deal team also received complete relationship briefs grounded in the firm’s records.
100+
portfolio evaluations with no hallucinations.
Each covered all 14 companies and eight quarters.
<60 sec
to map the strongest paths to a target.
The workflow resolves identities across the firm’s sources and returns recency and relationship-strength signals.
$0.20
in model input cost for one run of both workflows.
The same pair would cost an estimated $48.35 when intermediate data passes through the model.
Less token waste improved the answers.
Port of Context checks and organizes source data before the model responds. That keeps large volumes of intermediate material out of working context, so the model can finish from verified evidence without becoming overwhelmed, stopping early, or inventing details.
What Code Mode changed
Both agents finished the full job with far less model context.
The architecture behind both workflows is Code Mode. Without it, the model makes one search, reads the raw results, decides what to search next, and repeats, carrying every batch of data into the next step. With it, the model writes one program that runs the entire research sequence. The program searches, checks, and organizes the data, then returns the finished result: the ranked introduction path, or the complete portfolio view.
Network intelligence: six targets, one model turn
The program checked all six companies, tried another search when a match was unclear, traced the possible introduction paths, and ranked the results. The model then used that complete, ordered list to recommend the strongest route.
Code Mode, measured
11.4K
input tokens to resolve six companies and return the strongest relationship paths, in one model turn.
Conventional path, estimated
~131K
input tokens for the same recommendation across 12 model turns, with searches and candidate lists passing through the model.
91% fewer input tokens while returning the strongest introduction path.
Portfolio monitoring: 14 companies, eight quarters, one model turn
For each company, the financial data sits across multiple pages, and the update needs eight quarters in the same format. Code Mode let the agent bring together the complete picture for all 14 companies and turn it into one eight-quarter view.
Code Mode, measured
1.85K
input tokens to assemble the complete eight-quarter view, in one model turn.
Conventional path, estimated
~3.092M
input tokens for the same update across 84 model turns, with every page of financial data passing through the model.
99.9% fewer input tokens while returning the complete eight-quarter view.
Why it held up in production
Port of Context · Private Equity Case Study
