What clients say about working with us

Challenge

A freight company had attempted to apply a standard RL formulation to their vehicle routing problem but found that training agents on a simplified simulation produced policies that failed in production. The simulation didn't reflect real operational constraints — time windows, vehicle heterogeneity, driver shift limits.

Approach

Cognoscent rebuilt the simulation from the ground up, beginning with a constraint mapping exercise involving operations staff. The reward function was redesigned to balance three competing objectives rather than a single metric. Validation involved running historical routes through the trained policy and comparing against human dispatcher decisions.

Outcomes

The resulting environment supported agent training to a policy that performed comparably to senior dispatcher decisions on historical data. The client team continued using the environment for internal experimentation for at least four months post-engagement. Engagement ran 11 weeks against an 8–12 week scope.

Challenge

An investment team was evaluating an acquisition of a Singapore-based company whose primary asset was an AI-driven dynamic pricing engine. The company's technical documentation was substantial but the investment team lacked the internal capability to assess whether the claims about model performance were defensible.

Approach

The assessment covered model architecture, the data pipeline feeding the pricing engine, evaluation methodology (the company's own performance claims were based on backtesting that had several methodological issues), the engineering team's structure, and IP ownership. A three-week engagement with access to technical documentation and two structured interviews with the target company's engineering team.

Outcomes

The report identified that performance claims were based on a backtesting methodology that overstated out-of-sample performance. The data pipeline had two points of fragility that would require remediation before scaling. These findings were presented with severity ratings and mitigation options. The investment team used the report to negotiate price and agree remediation milestones in the acquisition agreement.

Challenge

A mid-sized professional services firm was losing senior practitioners to retirement, and with them a significant body of client engagement knowledge that existed only in personal drives, email archives, and project folders dating back twelve years.

Approach

The engagement began with a source inventory across approximately 40,000 documents across three storage systems. Extraction model design identified six primary entity types relevant to the firm's work. The extraction pipeline processed documents in batches with human review of sampled outputs at each stage. Quality metrics tracked entity completeness and relationship accuracy.

Outcomes

The structured output covered approximately 82% of accessible documents, with quality metrics showing high entity accuracy on modern documents and reduced accuracy on documents from before 2015. The pipeline documentation enabled the firm's IT team to run incremental extractions on new documents without further engagement. Delivered in five weeks.